Have you ever wondered why your home-brewed coffee never tastes as good as the coffee at your favorite coffee shop? Small changes in water temperature, grind size, water-to-coffee ratio, and brew time can have big effects on brew quality, which is why a systematic, scientific process is necessary for achieving the "perfect cup." In this presentation, we apply the design of experiments (DOE) methodology to the cold brew coffee extraction process. We explore process inputs, create and evaluate an experimental design, analyze the results, and build a statistical model to find the optimal cup. Attendees leave with a clear workflow for applying DOE in JMP, and, of course, the statistically backed recipe for the perfect cold brew!

 

 

Hello, everyone. My name is Rebecca Breuer, and I'm a statistician at Francis Medical. Today, I will actually be presenting a fun little side project that I did using DOE and JMP to optimize my cold brew coffee recipe. The plan for today is I will share a little bit about the basics of cold brew and the steps of the DOE methodology, and then we'll put the two together and talk about how to apply DOE and use JMP to optimize the cold brew coffee process.

First, a little bit about cold brew. Cold brew is a method of brewing coffee where you take some ground coffee and you combine it with some water. Usually the water is room temperature or cool or cold water, and then you let that mixture sit together in the refrigerator or maybe on the counter for some amount of time. Then you filter it. Finally, you serve it. Depending on how strong it is, you could drink it just as is, or you could dilute it with some water or ice or some milk.

I guess this process looks simple enough, right? Well, not so fast. Brewing coffee is actually quite a complex scientific process. Throughout each of those five steps that I mentioned before, there are several variables that can impact how your resulting brew turns out. I've listed several of them here, but this is by no means an exhaustive list. These are just some of the most important ones. I'll mention a few of the most important ones.

Obviously, the type of bean that you choose affects the final result, and how finely or coarsely you grind the beans, and then how much water you use compared to how much coffee you use. The water quality and composition is a big one. Then throughout the process, like when you're combining your ingredients and steeping, the amount of agitation or stirring matters, and also the temperature. Then, when you're letting your mixture steep, the amount of time that you let the mixture steep for matters. Then, when you're filtering, there are several things to consider such as the material of your filter, the size of the filter, how many times you filter it.

Then finally, when you serve it, if you add anything to your coffee, like if you dilute it or you add any flavors, obviously that changes the taste of it. But even things like the shape of your glassware could affect the flavor. There's a lot that goes into coffee. It's quite a complex process.

You might be wondering, how do we even quantify a resulting brew? How do we determine if it's good or not? To do that, there are two… We want to do that quantitatively and not just based on like, "Oh, I like this one. I don't like this one." There are two quantities that we can measure: total dissolved solids, or TDS, and extraction yield, or EY.

Total dissolved solids essentially quantifies the concentration or strength of your brew, and we can calculate that by measuring the mass of the coffee that's dissolved in our brew and dividing that by the overall mass of your final brew. If you have more solids dissolved in your coffee, then you'll have a stronger brew. If you have fewer solids dissolved in the water, then the brew will be weaker.

The second measure is extraction yield. Extraction yield is essentially the flavor profile of your coffee. Starting on the low end of extraction yield, we get the acidic and sour taste from the coffee. Those are the flavors that extract first. Then as we keep extracting a little longer, eventually we hit the ideal sweet spot where the coffee tastes sweet and smooth. Then if we keep extracting and then get high extraction yields, then our coffee starts to taste bitter and woody. We can calculate extraction yield, again, using that mass of solids that's dissolved in the water, and we divide that by the mass of the ground coffee that we put into our brew.

When we put those two measures together, we get this picture here, this little graph on the right side of the screen. This is a coffee control chart. It describes the flavor and the strength profile of different coffees. You can see that total dissolved solids is on the y-axis, and extraction yield is on the x-axis. I've got this area here that's highlighted in the dotted box, and that's the area that we're aiming for with our cold brew. I want a sweet, smooth-tasting cold brew. I don't know about you guys, but I don't think acidic, sour, bitter, woody, sounds very good.

Then for the total dissolved solids, since this is cold brew, I prefer to dilute my cold brew down with some milk or a little bit of cream at the end, so I want my brew to be more on the concentrate side. More concentrated would be better, but if it's a little less concentrated, that's okay with me. This is where we're aiming for: maximized total dissolved solids and an extraction yield between 18-22. That's what the literature suggests is ideal. But sometimes with cold brew, we might not quite get to 18, and we'll still get a sweet, smooth-tasting brew just because of how coffee flavors extract differently at colder temperatures. This is our sweet spot. This is our goal.

Now you're familiar with cold brew and what we're trying to do here, how we're trying to optimize the process. Let's talk a little bit about the DOE methodology. The DOE methodology is a structured approach for planning, conducting, analyzing, and interpreting an experiment. I've listed the steps out here in a little flowchart.

Basically, at the beginning of this process, we want to do some planning. We want to define the questions, define the problem that we're trying to solve, define the questions that we're trying to answer, and any objectives that we have for the experiment.

Then we want to identify our responses or what we're trying to measure, and also identify all the sources of variation that might affect your response. Also identify the ones that we think are most important and that we want to actively change in the experiment, and identify which ones we want to hold constant.

Then in the specify step, we want to take a first stab at like which statistical model are we trying to create? What kind of information do we want to get from our model? Do we want main effects? Do we need to be able to estimate curvature? We would specify that up front in the beginning.

Oftentimes, it can be helpful to run a pilot experiment in the beginning, and that helps you gather information to better inform your experiment. Then once we do some of this planning, then we can move on to actually designing the experiment.

Once we design the experiment and verify that it's going to give us the best information to answer our questions, then we'll go out and we will collect data, run the experiment, and then analyze our results in JMP, build a statistical model, and interpret it and ask ourselves if the objectives of our experiment were met and if we were able to answer the questions that we intended to. If we do, that's fantastic, and then we can report on our results, and we're done. But if we don't, we can always iterate. I think that's what's really cool about the DOE process: you can use one experiment to inform the next experiment.

Now we know about cold brew, and we know the DOE process, so let's put those two together and talk about how we can apply it to the cold brew coffee process. I'm going to walk through each of those steps from the DOE process, and we're going to lay out the framework for this specific problem. We'll also do some live JMP demo time to show how you can use JMP in these various steps.

First of all, I need to define my problem. My problem is that I want to make the best coffee or the best cold brew, and we're defining best by maximizing total dissolved solids and aiming for an extraction yield range of 18-22%. Also, I have some constraints where I only want to use equipment that I already have available at home, like my coffee scale, my Brita filter, my oven. I don't want to go out and buy expensive, specialized, multi-thousand-dollar equipment.

Also, I only want to actively vary three variables: grind size, water-to-coffee ratio, and the steep time. Then there's several other things that I'm going to hold constant, but these are the three factors that I think I can realistically vary in my process. The question I want to answer is: is there a combination of these key process variables that will meet my total dissolved solids and extraction yield goals?

Moving on to the identify step. I already mentioned that these are the three factors that I want to change. If you think back to one of my first slides about the cold brew process and I had all of those factors listed, these are the three that research suggests are most important and will have the biggest effect on my brew.

Grind size is the size of the coffee particle. I have a hand grinder that controls the grind size based on clicks, and the manufacturer of the grinder recommends a certain range of clicks for cold brew, and so I decided to take that range and go a little bit outside both ends. That was my range for the grind size factor. Second factor is the water-to-coffee ratio. For example, if I use 10 grams of coffee for one trial, then I would use 40 grams of water. Next is the steep time, and I'm going to vary that between 12 and 24 hours.

Then we already talked about the responses of total dissolved solids, the extraction yields, and our goals for both of those, so I won't repeat that.

One other thing that I want to mention in this step is I need to identify my resources. For this whole experiment, I went to my local grocery store, and I bought 500 grams of coffee beans. Assuming that I use 10 grams of coffee per run, then I have enough materials available for 50 runs.

Next, I want to specify the statistical model that I want to build and estimate. Since we have two responses, we're going to have two equations. I'm going to use linear regression, a linear regression model that includes main effects, quadratic effects, and the two-way interactions. That's based on research that I've done about cold brew and just coffee kinetics. This is the model that I think I need. In order to estimate all of these beta parameters, I'm going to need at least 10 runs, but more than that would be preferred because then I have degrees of freedom for error.

Let's go over to JMP now and talk about how, with all of these inputs in mind that we just discussed, let's see how we can design this experiment in JMP. I'm going to be using the custom design, so in the DOE tab, Custom Design. The first thing we want to do is we want to enter our responses and our factors, and I'm going to do that by using response and factor tables. I have this table set up right here that has my responses in it and the limits and then the goals for each one.

Then I've listed these two factors as being equally important, but you could change this. If you cared about extraction yield and the flavor more than you did the concentration, you could increase the importance of EY relative to TDS.

Then I have a similar table for my factors. I have each of my three factors here and the three levels that I'm interested in using in my experiment. Then, if we go back to the Custom Design tab in the red triangle, we can load the responses. I click on my response table and load it, and then I do the same thing for my factor table and load them.

I really like these response and factor tables because it makes creating DOEs just so easy. You don't have to type in each factor and the role and values every time. You can just load it. This is especially nice if maybe you're doing a pilot experiment where you're doing multiple DOEs, and let's say you do a pilot experiment and you find out that your range isn't large enough, you can just go into your table and you can change this value and then re-upload it. It makes it so easy, and it's a lot less typing.

You might have noticed that once I entered in my responses and my factors, JMP populated… It added some effects in this model tab. This is where we're going to specify the model that we're interested in estimating, and it's added the main effects by default. But we want to add a few more effects. I wanted to add the quadratic terms, and I also want to add my two-way interactions. It looks like we got everything in here now. I also want to make sure that where it says estimability, that everything is set to necessary, because I want to be able to estimate all of these.

Once we specify our model, then we can go down to number of runs, and the minimum is 10, but I'd like to do more than 10. I'm going to choose the default, which is 16. Then you would click Make design, but I already made my design, so I'm just going to go back to my JMP table where I have my DOE and my data, and we will look at this evaluate design tab.

Once you design a DOE, it's important to evaluate it and make sure that this DOE is going to give you the most information possible. This design evaluation tab in JMP has all those tools that can help you determine if your DOE is going to help meet your goals.

Some of the most important ones, I think, are the power analysis. We want to be able to detect these effects if they actually exist, so we're going to look for a power that's closer to one. For this particular experiment, it looks like we're doing pretty good with our main effects because we have a power of at least 80. Maybe not so good for some of these other terms down here, but I think I'm going to go with it because I can iterate later.

Also, if you have already done a previous pilot experiment and you have estimates of those beta coefficients and the root-mean-square error, you can change these values in here, and then it will update your power.

Another helpful one is the prediction variance profile, so I'll just take a look at this plot and looking for the variance to be pretty flat and even across my entire design space.

Then another one that I like is the color map on correlations. What we're looking for on this one is we want all the non-diagonal squares to be either white-colored or as light as possible. Looks like we got quite a bit of white in here and a little bit of light gray, so I'd say we're looking pretty good. I think we can use this design and go out and collect some data.

We have a design. Now we're ready to collect some data. You can see all the pictures from my process here, and it was quite the process. There were a lot of steps involved. I had to take measurements at several different points along my process and record those measurements, and then do some calculations to get my final total dissolved solids and extraction yield.

Basically, what I did was I ground my coffee, combined it, let it steep in the fridge, and then I did a double straining process because one strain with a mesh sieve wasn't enough. Then I had to evaporate my brew in the oven here to get the remaining dissolved solids that were in my mixture, because that's what we use to do our calculations. This is quite an involved process, and my advice here is this is why it can be really useful to run through your process ahead of time to make sure that the process is going to work for you. I went through several iterations of this process and did about three pilot studies in order to really nail down my process and figure out how I was going to measure and get the measurements that I needed.

After you run your experiment and you collect your data, now it's time to analyze. What I always like to do first is I like to look at my variables one at a time using the distribution platform. You can also use Graph Builder to explore variables two at a time or three at a time. Then we're going to use the Analyze tab in JMP to fit a model.

Let's go back over to JMP again, and we'll see how to do that. Here's my data table. You can see that I have my three inputs. Actually, there's two water-to-coffee ratios because I had target ratio, and then I have the actual ratio because I wasn't able to be perfect with my pouring and my measuring each time, so then I thought it would be better to use the actual ratio.

Then I have my two responses. Let's make some distributions of these. I'll highlight the ones that I want and go to the distribution platform and add them here and click OK. I'm just going to go and hover over each of the categories here from my inputs, and then we're going to look at how the response changes. It doesn't look like there's anything really obvious happening with total dissolved solids and extraction yield. Maybe some of the midpoint of grind size tends to be in the middle of both my outputs, but there's nothing standing out here.

Then I'm going to skip the water-to-coffee ratio for just a second, and we're going to look at steep time. Steep time, it doesn't really look like there's any patterns, either, with total dissolved solids or extraction yield. Nothing's standing out. But then if I go to water-to-coffee ratio, we'll click on the lowest ratio of about four, and this stands out. It looks like lower water-to-coffee ratio had higher total dissolved solids and lower extraction yield. Then middle values of our ratio correspond to the middle values of both of our outputs. Then the higher values of our ratio correspond to low values of TDS and high values of extraction yield.

With a little flavor of our data in mind, then we can build a model. We're going to use Fit Model. JMP actually will pre-populate the factors in the model that you specified in the beginning when you were designing the DOE, so it does this for you, which is really nice. I need to swap out this target water-to-coffee ratio with the actual. I already made this model, so we're going to choose all these settings, standard least squares, and just go to the model that I built. That's this one right here.

Next, we're going to take a look at this model, and we're going to evaluate it like you would go through all of your diagnostics and residual plots and look at things like how well the model fits and how much error you have. Let me just close this for a second. We have two tabs and two different equations to evaluate since we have our two outputs, so you want to check the residual plots for each one.

Then one thing that I want to highlight is this interactions profile plot. I think this is pretty cool. This plot is a way to graphically see any interactions between the factors. It's showing a red line and a blue line here for our low and high levels of our factors.

You can see that, just take grind size by water-to-coffee ratio, for example. We're looking at this box right here, and this is representing any possible interaction between these two factors. The lines look pretty parallel. They look pretty much the same. It's just the red one's higher than the blue one, so I'm not really anticipating any interaction between those two. But if we look at water-to-coffee ratio by steep time, it looks like the blue line starts on the bottom, and then as the ratio gets higher, the blue line goes up on top, above the red line. It looks like we might have an interaction there.

Then, of course, check your normal quantile plots for normality. This one looks pretty good. Our points follow the line decently, and they're within the dotted lines. We could do this same process of evaluating our model for the extraction response as well. I'm not going to go through all of that, but it looks good.

Then, once we are satisfied with this model, so just going back up to the top, these are all the terms that we are including in our model. If we wanted to, we could remove terms that are not significant and that are above a certain pre-specified p-value. But I think I'm just going to leave everything for now because based on theory, I know that there should be interactions and quadratic terms, so we're just going to leave it the way it is. Then once we're happy with our model, then we can actually predict.

We're going to use the prediction profiler, and we are going to try to maximize our goals. Just as a reminder, our goal for total dissolved solids is to maximize, and our goal for extraction yield is to match a target of 18-22. To do that, we can just hit Maximize desirability down here, and you can see that the profiler is going to change. It looks like JMP recommends This is our optimal cold brew recipe now that JMP is recommending. It's recommending to use a grind size of 25, a water-to-coffee ratio of about 11.7, and a steep time of 24 hours. Using those settings, we are expecting to get total dissolved solids of about 1.6% and extraction yield of about 16.4. This 16.4 does fall a little short of what we were aiming for, so maybe we have some work to do there. Maybe we could do another experiment.

Go back to the PowerPoint for a second, and we'll talk about that. We did our experiment. We analyzed it. We're interpreting it. Now let's report our results and make some conclusions.

As we saw in the Fit Model platform using the Prediction Profiler, a grind size of 25 clicks, a water-to-coffee ratio of 11.7, and a steep time of 24 hours resulted in the optimal cold brew. That had a total dissolved solids of about 1.6 and extraction yield of about 16.4. Like I said, we fell a little short of our extraction yield goal. Instead of doing another round of experiments right away, I wanted to taste it and see what it tasted like, because as I mentioned earlier, with cold brew, the extraction process is a little bit different than hot coffee because you're brewing at a colder temperature, so we can still have sweet, smooth coffee with an extraction yield a little bit below 18.

I wanted to taste and see. I made a couple of samples with these desired or optimal settings, and I tasted it, and they all tasted really good. I decided that I was happy, and I was going to keep using this recipe. But if I did want to keep going, we do have some potential next steps here. The first one is maybe we could actually create some more make some more samples and go through the whole process of measuring and putting the samples in the oven and performing the experiment again just to get a few confirmation runs that the EY and total dissolved TDS values are consistent for our optimal settings.

Another potential next step is augmenting the DOE to try to achieve a higher extraction yield. Maybe we could expand some of our factor ranges a little bit to try to push the limits of extraction yield.

Then the third potential next step is maybe we could look at using a more complicated model. Maybe it would be helpful to transform some of our variables or maybe look at using a non-linear model because those models might more accurately represent the coffee rate kinetics that are happening.

That's all I have for you today. Thank you for listening, and I hope that the next time you're enjoying your next cup of coffee, you really appreciate all the science that goes into making it taste great. Thank you.

0 Comments
Presented At Discovery Summit 2026

Presenter

Skill level

Intermediate
  • Beginner
  • Intermediate
  • Advanced

Files

Published on ‎07-15-2026 03:39 PM by Community Manager Community Manager | Updated on ‎10-30-2026 03:34 PM

Have you ever wondered why your home-brewed coffee never tastes as good as the coffee at your favorite coffee shop? Small changes in water temperature, grind size, water-to-coffee ratio, and brew time can have big effects on brew quality, which is why a systematic, scientific process is necessary for achieving the "perfect cup." In this presentation, we apply the design of experiments (DOE) methodology to the cold brew coffee extraction process. We explore process inputs, create and evaluate an experimental design, analyze the results, and build a statistical model to find the optimal cup. Attendees leave with a clear workflow for applying DOE in JMP, and, of course, the statistically backed recipe for the perfect cold brew!

 

 

Hello, everyone. My name is Rebecca Breuer, and I'm a statistician at Francis Medical. Today, I will actually be presenting a fun little side project that I did using DOE and JMP to optimize my cold brew coffee recipe. The plan for today is I will share a little bit about the basics of cold brew and the steps of the DOE methodology, and then we'll put the two together and talk about how to apply DOE and use JMP to optimize the cold brew coffee process.

First, a little bit about cold brew. Cold brew is a method of brewing coffee where you take some ground coffee and you combine it with some water. Usually the water is room temperature or cool or cold water, and then you let that mixture sit together in the refrigerator or maybe on the counter for some amount of time. Then you filter it. Finally, you serve it. Depending on how strong it is, you could drink it just as is, or you could dilute it with some water or ice or some milk.

I guess this process looks simple enough, right? Well, not so fast. Brewing coffee is actually quite a complex scientific process. Throughout each of those five steps that I mentioned before, there are several variables that can impact how your resulting brew turns out. I've listed several of them here, but this is by no means an exhaustive list. These are just some of the most important ones. I'll mention a few of the most important ones.

Obviously, the type of bean that you choose affects the final result, and how finely or coarsely you grind the beans, and then how much water you use compared to how much coffee you use. The water quality and composition is a big one. Then throughout the process, like when you're combining your ingredients and steeping, the amount of agitation or stirring matters, and also the temperature. Then, when you're letting your mixture steep, the amount of time that you let the mixture steep for matters. Then, when you're filtering, there are several things to consider such as the material of your filter, the size of the filter, how many times you filter it.

Then finally, when you serve it, if you add anything to your coffee, like if you dilute it or you add any flavors, obviously that changes the taste of it. But even things like the shape of your glassware could affect the flavor. There's a lot that goes into coffee. It's quite a complex process.

You might be wondering, how do we even quantify a resulting brew? How do we determine if it's good or not? To do that, there are two… We want to do that quantitatively and not just based on like, "Oh, I like this one. I don't like this one." There are two quantities that we can measure: total dissolved solids, or TDS, and extraction yield, or EY.

Total dissolved solids essentially quantifies the concentration or strength of your brew, and we can calculate that by measuring the mass of the coffee that's dissolved in our brew and dividing that by the overall mass of your final brew. If you have more solids dissolved in your coffee, then you'll have a stronger brew. If you have fewer solids dissolved in the water, then the brew will be weaker.

The second measure is extraction yield. Extraction yield is essentially the flavor profile of your coffee. Starting on the low end of extraction yield, we get the acidic and sour taste from the coffee. Those are the flavors that extract first. Then as we keep extracting a little longer, eventually we hit the ideal sweet spot where the coffee tastes sweet and smooth. Then if we keep extracting and then get high extraction yields, then our coffee starts to taste bitter and woody. We can calculate extraction yield, again, using that mass of solids that's dissolved in the water, and we divide that by the mass of the ground coffee that we put into our brew.

When we put those two measures together, we get this picture here, this little graph on the right side of the screen. This is a coffee control chart. It describes the flavor and the strength profile of different coffees. You can see that total dissolved solids is on the y-axis, and extraction yield is on the x-axis. I've got this area here that's highlighted in the dotted box, and that's the area that we're aiming for with our cold brew. I want a sweet, smooth-tasting cold brew. I don't know about you guys, but I don't think acidic, sour, bitter, woody, sounds very good.

Then for the total dissolved solids, since this is cold brew, I prefer to dilute my cold brew down with some milk or a little bit of cream at the end, so I want my brew to be more on the concentrate side. More concentrated would be better, but if it's a little less concentrated, that's okay with me. This is where we're aiming for: maximized total dissolved solids and an extraction yield between 18-22. That's what the literature suggests is ideal. But sometimes with cold brew, we might not quite get to 18, and we'll still get a sweet, smooth-tasting brew just because of how coffee flavors extract differently at colder temperatures. This is our sweet spot. This is our goal.

Now you're familiar with cold brew and what we're trying to do here, how we're trying to optimize the process. Let's talk a little bit about the DOE methodology. The DOE methodology is a structured approach for planning, conducting, analyzing, and interpreting an experiment. I've listed the steps out here in a little flowchart.

Basically, at the beginning of this process, we want to do some planning. We want to define the questions, define the problem that we're trying to solve, define the questions that we're trying to answer, and any objectives that we have for the experiment.

Then we want to identify our responses or what we're trying to measure, and also identify all the sources of variation that might affect your response. Also identify the ones that we think are most important and that we want to actively change in the experiment, and identify which ones we want to hold constant.

Then in the specify step, we want to take a first stab at like which statistical model are we trying to create? What kind of information do we want to get from our model? Do we want main effects? Do we need to be able to estimate curvature? We would specify that up front in the beginning.

Oftentimes, it can be helpful to run a pilot experiment in the beginning, and that helps you gather information to better inform your experiment. Then once we do some of this planning, then we can move on to actually designing the experiment.

Once we design the experiment and verify that it's going to give us the best information to answer our questions, then we'll go out and we will collect data, run the experiment, and then analyze our results in JMP, build a statistical model, and interpret it and ask ourselves if the objectives of our experiment were met and if we were able to answer the questions that we intended to. If we do, that's fantastic, and then we can report on our results, and we're done. But if we don't, we can always iterate. I think that's what's really cool about the DOE process: you can use one experiment to inform the next experiment.

Now we know about cold brew, and we know the DOE process, so let's put those two together and talk about how we can apply it to the cold brew coffee process. I'm going to walk through each of those steps from the DOE process, and we're going to lay out the framework for this specific problem. We'll also do some live JMP demo time to show how you can use JMP in these various steps.

First of all, I need to define my problem. My problem is that I want to make the best coffee or the best cold brew, and we're defining best by maximizing total dissolved solids and aiming for an extraction yield range of 18-22%. Also, I have some constraints where I only want to use equipment that I already have available at home, like my coffee scale, my Brita filter, my oven. I don't want to go out and buy expensive, specialized, multi-thousand-dollar equipment.

Also, I only want to actively vary three variables: grind size, water-to-coffee ratio, and the steep time. Then there's several other things that I'm going to hold constant, but these are the three factors that I think I can realistically vary in my process. The question I want to answer is: is there a combination of these key process variables that will meet my total dissolved solids and extraction yield goals?

Moving on to the identify step. I already mentioned that these are the three factors that I want to change. If you think back to one of my first slides about the cold brew process and I had all of those factors listed, these are the three that research suggests are most important and will have the biggest effect on my brew.

Grind size is the size of the coffee particle. I have a hand grinder that controls the grind size based on clicks, and the manufacturer of the grinder recommends a certain range of clicks for cold brew, and so I decided to take that range and go a little bit outside both ends. That was my range for the grind size factor. Second factor is the water-to-coffee ratio. For example, if I use 10 grams of coffee for one trial, then I would use 40 grams of water. Next is the steep time, and I'm going to vary that between 12 and 24 hours.

Then we already talked about the responses of total dissolved solids, the extraction yields, and our goals for both of those, so I won't repeat that.

One other thing that I want to mention in this step is I need to identify my resources. For this whole experiment, I went to my local grocery store, and I bought 500 grams of coffee beans. Assuming that I use 10 grams of coffee per run, then I have enough materials available for 50 runs.

Next, I want to specify the statistical model that I want to build and estimate. Since we have two responses, we're going to have two equations. I'm going to use linear regression, a linear regression model that includes main effects, quadratic effects, and the two-way interactions. That's based on research that I've done about cold brew and just coffee kinetics. This is the model that I think I need. In order to estimate all of these beta parameters, I'm going to need at least 10 runs, but more than that would be preferred because then I have degrees of freedom for error.

Let's go over to JMP now and talk about how, with all of these inputs in mind that we just discussed, let's see how we can design this experiment in JMP. I'm going to be using the custom design, so in the DOE tab, Custom Design. The first thing we want to do is we want to enter our responses and our factors, and I'm going to do that by using response and factor tables. I have this table set up right here that has my responses in it and the limits and then the goals for each one.

Then I've listed these two factors as being equally important, but you could change this. If you cared about extraction yield and the flavor more than you did the concentration, you could increase the importance of EY relative to TDS.

Then I have a similar table for my factors. I have each of my three factors here and the three levels that I'm interested in using in my experiment. Then, if we go back to the Custom Design tab in the red triangle, we can load the responses. I click on my response table and load it, and then I do the same thing for my factor table and load them.

I really like these response and factor tables because it makes creating DOEs just so easy. You don't have to type in each factor and the role and values every time. You can just load it. This is especially nice if maybe you're doing a pilot experiment where you're doing multiple DOEs, and let's say you do a pilot experiment and you find out that your range isn't large enough, you can just go into your table and you can change this value and then re-upload it. It makes it so easy, and it's a lot less typing.

You might have noticed that once I entered in my responses and my factors, JMP populated… It added some effects in this model tab. This is where we're going to specify the model that we're interested in estimating, and it's added the main effects by default. But we want to add a few more effects. I wanted to add the quadratic terms, and I also want to add my two-way interactions. It looks like we got everything in here now. I also want to make sure that where it says estimability, that everything is set to necessary, because I want to be able to estimate all of these.

Once we specify our model, then we can go down to number of runs, and the minimum is 10, but I'd like to do more than 10. I'm going to choose the default, which is 16. Then you would click Make design, but I already made my design, so I'm just going to go back to my JMP table where I have my DOE and my data, and we will look at this evaluate design tab.

Once you design a DOE, it's important to evaluate it and make sure that this DOE is going to give you the most information possible. This design evaluation tab in JMP has all those tools that can help you determine if your DOE is going to help meet your goals.

Some of the most important ones, I think, are the power analysis. We want to be able to detect these effects if they actually exist, so we're going to look for a power that's closer to one. For this particular experiment, it looks like we're doing pretty good with our main effects because we have a power of at least 80. Maybe not so good for some of these other terms down here, but I think I'm going to go with it because I can iterate later.

Also, if you have already done a previous pilot experiment and you have estimates of those beta coefficients and the root-mean-square error, you can change these values in here, and then it will update your power.

Another helpful one is the prediction variance profile, so I'll just take a look at this plot and looking for the variance to be pretty flat and even across my entire design space.

Then another one that I like is the color map on correlations. What we're looking for on this one is we want all the non-diagonal squares to be either white-colored or as light as possible. Looks like we got quite a bit of white in here and a little bit of light gray, so I'd say we're looking pretty good. I think we can use this design and go out and collect some data.

We have a design. Now we're ready to collect some data. You can see all the pictures from my process here, and it was quite the process. There were a lot of steps involved. I had to take measurements at several different points along my process and record those measurements, and then do some calculations to get my final total dissolved solids and extraction yield.

Basically, what I did was I ground my coffee, combined it, let it steep in the fridge, and then I did a double straining process because one strain with a mesh sieve wasn't enough. Then I had to evaporate my brew in the oven here to get the remaining dissolved solids that were in my mixture, because that's what we use to do our calculations. This is quite an involved process, and my advice here is this is why it can be really useful to run through your process ahead of time to make sure that the process is going to work for you. I went through several iterations of this process and did about three pilot studies in order to really nail down my process and figure out how I was going to measure and get the measurements that I needed.

After you run your experiment and you collect your data, now it's time to analyze. What I always like to do first is I like to look at my variables one at a time using the distribution platform. You can also use Graph Builder to explore variables two at a time or three at a time. Then we're going to use the Analyze tab in JMP to fit a model.

Let's go back over to JMP again, and we'll see how to do that. Here's my data table. You can see that I have my three inputs. Actually, there's two water-to-coffee ratios because I had target ratio, and then I have the actual ratio because I wasn't able to be perfect with my pouring and my measuring each time, so then I thought it would be better to use the actual ratio.

Then I have my two responses. Let's make some distributions of these. I'll highlight the ones that I want and go to the distribution platform and add them here and click OK. I'm just going to go and hover over each of the categories here from my inputs, and then we're going to look at how the response changes. It doesn't look like there's anything really obvious happening with total dissolved solids and extraction yield. Maybe some of the midpoint of grind size tends to be in the middle of both my outputs, but there's nothing standing out here.

Then I'm going to skip the water-to-coffee ratio for just a second, and we're going to look at steep time. Steep time, it doesn't really look like there's any patterns, either, with total dissolved solids or extraction yield. Nothing's standing out. But then if I go to water-to-coffee ratio, we'll click on the lowest ratio of about four, and this stands out. It looks like lower water-to-coffee ratio had higher total dissolved solids and lower extraction yield. Then middle values of our ratio correspond to the middle values of both of our outputs. Then the higher values of our ratio correspond to low values of TDS and high values of extraction yield.

With a little flavor of our data in mind, then we can build a model. We're going to use Fit Model. JMP actually will pre-populate the factors in the model that you specified in the beginning when you were designing the DOE, so it does this for you, which is really nice. I need to swap out this target water-to-coffee ratio with the actual. I already made this model, so we're going to choose all these settings, standard least squares, and just go to the model that I built. That's this one right here.

Next, we're going to take a look at this model, and we're going to evaluate it like you would go through all of your diagnostics and residual plots and look at things like how well the model fits and how much error you have. Let me just close this for a second. We have two tabs and two different equations to evaluate since we have our two outputs, so you want to check the residual plots for each one.

Then one thing that I want to highlight is this interactions profile plot. I think this is pretty cool. This plot is a way to graphically see any interactions between the factors. It's showing a red line and a blue line here for our low and high levels of our factors.

You can see that, just take grind size by water-to-coffee ratio, for example. We're looking at this box right here, and this is representing any possible interaction between these two factors. The lines look pretty parallel. They look pretty much the same. It's just the red one's higher than the blue one, so I'm not really anticipating any interaction between those two. But if we look at water-to-coffee ratio by steep time, it looks like the blue line starts on the bottom, and then as the ratio gets higher, the blue line goes up on top, above the red line. It looks like we might have an interaction there.

Then, of course, check your normal quantile plots for normality. This one looks pretty good. Our points follow the line decently, and they're within the dotted lines. We could do this same process of evaluating our model for the extraction response as well. I'm not going to go through all of that, but it looks good.

Then, once we are satisfied with this model, so just going back up to the top, these are all the terms that we are including in our model. If we wanted to, we could remove terms that are not significant and that are above a certain pre-specified p-value. But I think I'm just going to leave everything for now because based on theory, I know that there should be interactions and quadratic terms, so we're just going to leave it the way it is. Then once we're happy with our model, then we can actually predict.

We're going to use the prediction profiler, and we are going to try to maximize our goals. Just as a reminder, our goal for total dissolved solids is to maximize, and our goal for extraction yield is to match a target of 18-22. To do that, we can just hit Maximize desirability down here, and you can see that the profiler is going to change. It looks like JMP recommends This is our optimal cold brew recipe now that JMP is recommending. It's recommending to use a grind size of 25, a water-to-coffee ratio of about 11.7, and a steep time of 24 hours. Using those settings, we are expecting to get total dissolved solids of about 1.6% and extraction yield of about 16.4. This 16.4 does fall a little short of what we were aiming for, so maybe we have some work to do there. Maybe we could do another experiment.

Go back to the PowerPoint for a second, and we'll talk about that. We did our experiment. We analyzed it. We're interpreting it. Now let's report our results and make some conclusions.

As we saw in the Fit Model platform using the Prediction Profiler, a grind size of 25 clicks, a water-to-coffee ratio of 11.7, and a steep time of 24 hours resulted in the optimal cold brew. That had a total dissolved solids of about 1.6 and extraction yield of about 16.4. Like I said, we fell a little short of our extraction yield goal. Instead of doing another round of experiments right away, I wanted to taste it and see what it tasted like, because as I mentioned earlier, with cold brew, the extraction process is a little bit different than hot coffee because you're brewing at a colder temperature, so we can still have sweet, smooth coffee with an extraction yield a little bit below 18.

I wanted to taste and see. I made a couple of samples with these desired or optimal settings, and I tasted it, and they all tasted really good. I decided that I was happy, and I was going to keep using this recipe. But if I did want to keep going, we do have some potential next steps here. The first one is maybe we could actually create some more make some more samples and go through the whole process of measuring and putting the samples in the oven and performing the experiment again just to get a few confirmation runs that the EY and total dissolved TDS values are consistent for our optimal settings.

Another potential next step is augmenting the DOE to try to achieve a higher extraction yield. Maybe we could expand some of our factor ranges a little bit to try to push the limits of extraction yield.

Then the third potential next step is maybe we could look at using a more complicated model. Maybe it would be helpful to transform some of our variables or maybe look at using a non-linear model because those models might more accurately represent the coffee rate kinetics that are happening.

That's all I have for you today. Thank you for listening, and I hope that the next time you're enjoying your next cup of coffee, you really appreciate all the science that goes into making it taste great. Thank you.



Start:
Wed, Oct 21, 2026 11:30 AM EDT
End:
Wed, Oct 21, 2026 12:15 PM EDT
Key Ballroom 9
Attachments
0 Kudos