Starter kit for photovoltaics optimization using machine learning.
This kit includes useful data visualization tools as well as a starter package for Bayesian optimization of photovoltaics data. The code to run this starter kit is found in the file Run_Starter_Kit.ipynb
The input data consists of a .csv file with each experiment formatted into rows with each column being a real, continuous variable. The last column of the file should be the target variable that will be optimized. An example dataset on how to properly format the file is found here [1].
All required packages and the corresponding versions can be found in the requirements.txt file. Once installed in the Python environment of choice, the code can be run.
Several forms of data visualization are used to help the user better understand the biases, correlations, and variable importance within the dataset [2].
- 1D and 2D Histograms plot the distribution of data against another variable.
- Correlation Coefficient Matrices show how closely any two variables are coupled to each other.
- SHAP Feature Importance ranks each variable as how important it is in determining the magnitude and direction of the target output variable values.
Bayesian optimization (BO) stems from the principles of Bayes' Theorem. BO is a probabilistic tool used to help guide the selection of new data points within a dataset based on the computed means and variances of a surrogate model. BO is useful for driving experimental optimization when data are scarce or expensive to obtain.
By running the code found in Run_Starter_Kit.ipynb, a set of suggested conditions will be presented by the algorithm. These conditions either exploit the predicted means to obtain a highly optimum sample or explore the high variance regions of the dataset to minimize uncertainity for the next round of sampling.
[1] Liu, Z. et al. Machine learning with knowledge constraints for process optimization of open-air perovskite solar cell manufacturing. Joule 6, 834–849 (2022).
[2] MIT Subject 2.S986, Fall 2020 GitHub.