Experiment Versus Computation, Physics Versus Machine Learning in The Inverse Design of Photovoltaic Materials

As a practical case study to advance methods for the inverse design of photovoltaic (PV) materials, my group has been given the following task for the next 5 years: Identify and synthesize the best PV absorber out of the space of all inorganic materials containing sulfur and phosphorus (phosphosulfides). Within this family of materials ~10 compounds have already been considered as potential PV materials, ~260 have previously been synthesized, ~420 are present in the Materials Project database, and 200,000 phosphosulfide compounds from ternary and quaternary systems are chemically plausible according to charge-neutrality arguments.<br/><br/>To try to solve this puzzle within a 5-year timeframe, we have set up a multi-faceted inverse design platform. This platform includes:<br/><br/>1. High-throughput/high versatility modular synthesis apparatus for combinatorial growth of any inorganic phosphosulfide thin films, so that there are no excuses for excluding certain elements from experimental work. This includes sulfur- and phosphorus partial pressure control (both at atmospheric and low total pressures), separate chambers for volatile metal incorporation, and creation of perpendicular combinatorial gradients in the metal-to-metal ratio as well as in the S/P ratio.<br/><br/>2. High-throughput characterization apparatus, focusing on the properties that are expected to be correlated with the final PV performance and are difficult to simulate with sufficient accuracy.<br/><br/>3. A tiered approach to material property simulation, with a focus on complementing (rather than reiterating) the information available from experiment. The simulation tools can be roughly divided into first-principles quantum mechanical methods, semiclassical methods, and classical “rule-of-thumb“ methods.<br/><br/>4. A data management tool to accommodate both experimental and simulation data.<br/><br/>5. Various artificial intelligence tools to be used both for decision-making in the lab and for understanding complex composition-structure-process-property-performance relationships.<br/><br/>6. Public access to the data generated by the tools listed above.<br/><br/>To make the presentation more concrete, I will discuss two practical examples that highlight the importance of a diversified methodological toolkit in modern materials research. The first example is the integration of computational and experimental approaches in the process of discovering new materials with new compositions and unknown crystal structures. The second example is the balance between physical soundness and statistical performance in the machine-learning-aided development of a phenomenological figure of merit to evaluate the quality of a generic PV material.