[Paper Review] Embedding Physics Domain Knowledge into a Bayesian Network Enables Layer-by-Layer Process Innovation for Photovoltaics

Zekun Ren, Felipe Oviedo+16 more|arXiv (Cornell University)|Jul 25, 2019

solar cell performance optimization46 references4 citations

TL;DR

This paper proposes a physics-informed Bayesian network that integrates domain knowledge with a deep autoencoder-based surrogate model to enable layer-by-layer optimization of GaAs solar cell processes. By linking process parameters to materials properties and device performance, the method achieves a 6.5% relative efficiency improvement over baseline and black-box methods with minimal experimental iteration, demonstrating transparent, causal diagnosis of performance limitations at the layer level.

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 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 layer-by-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.

Motivation & Objective

• To overcome the limitations of trial-and-error, black-box optimization in photovoltaic process development, which often lacks transparency and global optimality.
• To integrate physics-based domain knowledge—such as bulk lifetime, doping, and surface recombination—into a probabilistic framework for causal inference.
• To reduce experimental burden by enabling accurate performance prediction with few fabricated samples through a fast surrogate model.
• To achieve layer-by-layer process innovation by identifying root causes of underperformance and revealing alternative optimal process windows.

Proposed method

• A Bayesian network is constructed to model causal relationships between process conditions, materials descriptors (e.g., bulk lifetime, doping, surface recombination), and device performance (e.g., cell efficiency).
• An autoencoder-based surrogate device-physics model is trained to predict device performance up to 100× faster than traditional numerical solvers.
• Physics domain knowledge is embedded into the Bayesian network structure and conditional probability distributions to ensure physically plausible inferences.
• Bayesian inference is used to update beliefs about process parameters based on limited experimental data, enabling efficient learning with small datasets.
• The framework supports layer-wise diagnosis by isolating the impact of individual processing steps on overall device performance.
• The method is validated through a single-cycle optimization of GaAs solar cells, identifying an improved growth temperature without secondary measurements.

Experimental results

Research questions

• RQ1 Can physics domain knowledge be effectively embedded into a Bayesian network to improve interpretability and accuracy in photovoltaic process optimization?
• RQ2 How does the integration of a fast surrogate model enhance the efficiency of Bayesian inference in device process development?
• RQ3 To what extent can layer-by-layer diagnosis identify the root causes of performance degradation in multilayer photovoltaic devices?
• RQ4 Does the proposed method outperform traditional black-box optimization in terms of efficiency gain and experimental efficiency?
• RQ5 Can the framework discover alternative optimal process windows beyond those identified by standard optimization approaches?

Key findings

• The method achieved a 6.5% relative improvement in AM1.5G power conversion efficiency over the baseline and traditional black-box optimization methods.
• The framework enabled identification of an improved growth temperature for GaAs solar cells without requiring additional experimental measurements.
• The surrogate model reduced prediction time by a factor of 100 compared to numerical solvers, enabling rapid iteration with minimal experimental input.
• The Bayesian network provided transparent, causal insights into performance limitations at the layer level, revealing specific process conditions responsible for underperformance.
• The approach demonstrated robustness and high sample efficiency, achieving significant optimization gains with only a small number of fabricated samples.
• The method uncovered alternative optimal process windows that were not identified by global black-box optimization, highlighting its ability to escape local optima.