Enabling high-throughput electrical conductivity optimization of doped conjugated polymers using explainable machine learning

The recent advent of workflows involving high-throughput experimentation techniques, in combination with machine learning optimization, has enabled the accelerated discovery of materials with state-of-the-art properties. However, there remains many other workflows which require measurements of quantities that cannot be easily automated, thereby limiting discovery. In particular, the optimization of the electrical conductivity of doped polymer materials requires laborious measurements. Here, we propose a workflow which involves a data-driven approach using a pair of classification and regression models to reduce the need for manual intervention. The first model classifies the samples at an accuracy of 100% for conductivity >~ 25 to 100 S/cm. We predicted the conductivities of the smaller subset of samples using a second model, which has test R ² of 0.984. To validate the approach, we showed that the models, neither trained on the samples with the two highest conductivities of 498 and 506 S/cm, were able to, in an extrapolative manner, correctly classify and predict them at errors of 137 and 140 S/cm, respectively. The workflow results in an improvement in measurement efficiency by 59%, which is 89% of the maximum achievable. Concurrently, we addressed the typical lack of explainability of machine-learned models by engineering bespoke features and coupling them with carefully selected models, and derived insights on the spectral influence of conductivity. This study offers a way to accelerate the optimization of doped polymer materials and demonstrates the insights that can be gained with deliberate uses of machine learning in experimental science.