Embedding physics domain knowledge into a Bayesian network enables layer-by-layer process innovation for photovoltaics

Abstract

Process optimization of photovoltaic devices is a time-intensive, trial-and-error endeavor, without full transparency of the underlying physics, and with user-imposed constraints that may or may not lead to a global optimum. Herein, we demonstrate that embedding physics domain knowledge into a Bayesian network enables an optimization approach that identifies the root cause(s) of underperformance with layer-by-layer resolution and reveals alternative optimal process windows beyond global black-box 2 optimization. Our Bayesian-network approach links process conditions to materials descriptors (bulk and interface properties, e.g., bulk lifetime, doping, and surface recombination) and device-performance parameters (e.g., cell efficiency), using a Bayesian inference framework with an autoencoder-based surrogate device-physics model that is 100x faster than numerical solvers. With the trained surrogate model, our approach is robust and reduces significantly the time-consuming experimentalist intervention, even with small numbers of fabricated samples. To demonstrate our method, we perform layerby-layer optimization of GaAs solar cells. In a single cycle of learning, we find an improved growth temperature for the GaAs solar cells without any secondary measurements, and demonstrate a 6.5% relative AM1.5G efficiency improvement above baseline and traditional black-box optimization methods.

Toward autonomous design and synthesis of novel inorganic materials

“…Modified outcomes Some common sources and representations of information include physics-based relations described by algebraic or differential equations, probabilistic relationships derived from previous experimental or theoretical findings, and logic-based rules or graphs developed using predictions from first principles calculations or expert intuition. To integrate this information into the optimization process, the selection of experiments can be biased toward subspaces where the objective is expected to be optimal 134 , positive (negative) weights can be placed on particularly (un)desirable outcomes 135 , and results can be validated through a comparison with known laws or empirical relationships 136 . In each case, all information can be incorporated prior to beginning the experimental procedure via data fusion.…”

AI Applications through the Whole Life Cycle of Material Discovery

“…For example, ML methods have already managed to (1) automate materials' characterization processes and effectively analyze the characterization dataset, [18][19][20][21] (2) quickly screen the vast material design space (e.g., reducing the prediction time of DFT from 10 3 s to 10 À2 s), [22][23][24][25] (3) realize property prediction in complex material systems with limited first-principles understanding, 26 (4) directly map high-dimensional synthesis recipes to materials with desired properties, 27,28 and (5) extract generalizable scientific principles from various material systems. 27,29,30 The reason why AI is particularly apt in material design is due to its inherently strong capabilities in handling huge amounts of data as well as high-dimensional analysis. A single material type within a synthesis protocol could contain enormous intrinsic information, such as various physiochemical properties, chemical structure, and composition information.…”

“…The two-step highly interpretable Bayesian inference framework is shown in Figure 9A. 30 New physical interpretations are also generated for complex, high-dimensional grain-boundary systems. 150 Several ML methods are combined with a highquality structure descriptor called smooth overlap of atomic positions (SOAP) to capture local atomic environment (LAE) in predicting several crystalline properties.…”

“…Instead of mapping the relations between process variables and device performance of photovoltaics directly, a two-step ML framework is proposed to connect target variables to material descriptors first, then to process conditions. 30 Such transparency can enable a better understanding of the whole prediction workflow, which in turn gives more interpretability. For different material systems, the detailed decompositions of tasks are different; however, there will be several general frameworks that could guide different material systems.…”

“…AI-Aided Theory Paradigm Discovery (A) Schematic of an interpretable ML framework as a physically informed two-step Bayesian network-based process-optimization model. It first links process conditions to material descriptors, then the latter to device performance 30. (B) Graphical representation of the three hypotheses generated from the model, and representative structures for each hypothesis.…”