Machine Learning for your Lab
Have you ever spent hours on your phone, watching short videos that never seem to end? You ask yourself why you stayed hooked for so long. The algorithms that pick the videos wanted exactly that: your attention, for as long as possible. It watches how you behave, then serves the next video you cannot resist.
These are powerful algorithms. They show you a video, measure how you react to it, then predict the next that will keep you watching. Every video is an experiment. The algorithm runs it, reads the outcome, then runs another one that raises the odds of keeping you on the platform.
What if your lab had something similar? An algorithm that reads the outcome of each experiment and points you toward the settings that improve your results. Alfraido does exactly that.
Like the algorithms behind YouTube and Meta, Alfraido is seamless. It works under the hood. It fits into a normal lab workflow without changing how you work.
Formulators are not Machine Learning Engineers
Data science and machine learning are powerful fields. Data scientists explore data, find the underlying trends, and build models that link target variables to explanatory variables. From those models, they predict outcomes before running experiments.
Machine learning is useful in any lab. Many labs already use it through various tools. But these tools speak a language built for data scientists, not for materials formulators. A formulator should not have to become a data scientist to harness the full power of machine learning. You should not need to know what XGBoost, a neural network, or a random forest is to get results.
Alfraido's workflow
Alfraido follows a simple workflow.
Creating a project
When you create a project, you tell Alfraido which outcomes you want to maximize, minimize, or hold within a limit. You might for example maximize strength while minimizing co2_footprint.
You can also keep other outcomes under control (eg. slump >= 200) while you optimize your objectives.
Next, you tell Alfraido which parameters you control and how much room you have to change each one. These are things like ingredient dosages, production temperature, pressure, and the type of raw material.
You can add constraints on parameters. For example, your ingredient dosages might need to sum to 100%.
Finally, you define the team that will work on the project.
Collaborative data logging
Once the project is set up, you upload your existing experiment results as a CSV table.
The model reads the data and tells you what to test next, along with its prediction of the outcome.
You assign the proposed experiments to your team as tasks. You can also design an experiment yourself.
Team members get an email when they have new tasks. They run the experiments and submit the results.
When the manager validates a result, it feeds back into the model, which then points to where to look next.
Assess the quality of the predictive model
You measure a model's quality by its accuracy. You hide one experiment from the model, ask it to predict that experiment's outcome, then compare the prediction to the outcome you actually observed.
Do this for every experiment and you get a cross validation plot. The closer the points sit to the diagonal, the more accurate the model.
Alfraido keeps this plot on the project dashboard at all times. You can see when the model is making useful predictions and when it still needs more data.
Track your progress
When you optimize a single metric, such as strength, tracking progress is simple. You watch how much the objective improves.
When you optimize several objectives at once, such as maximizing strength while minimizing co2, you track progress through the Pareto front.
The Pareto front is the curve, surface, or hypersurface that holds the best tradeoffs you have found so far. Pushing the optimization forward means pushing the Pareto front outward. The area, volume, or hypervolume under it grows. That area under the Pareto front is your main progress metric in this case.
