[Paper Review] Interpretable deep learning for guided structure-property explorations in photovoltaics

Balaji Sesha Sarath Pokuri, Sambuddha Ghosal+3 more|arXiv (Cornell University)|Nov 14, 2018

Machine Learning in Materials ScienceMaterials Science17 references3 citations

TL;DR

This paper proposes a custom deep convolutional neural network (CNN) as a fast, interpretable surrogate model to map organic photovoltaic (OPV) active layer microstructures to their power conversion efficiency, enabling guided morphology design. The model achieves high accuracy in classifying performance levels and identifies critical interfacial features through saliency maps, outperforming standard networks like VGG-16 and ResNet-50 in interpretability and generalization on out-of-sample data.

ABSTRACT

The performance of an organic photovoltaic device is intricately connected to its active layer morphology. This connection between the active layer and device performance is very expensive to evaluate, either experimentally or computationally. Hence, designing morphologies to achieve higher performances is non-trivial and often intractable. To solve this, we first introduce a deep convolutional neural network (CNN) architecture that can serve as a fast and robust surrogate for the complex structure-property map. Several tests were performed to gain trust in this trained model. Then, we utilize this fast framework to perform robust microstructural design to enhance device performance.

Motivation & Objective

• Address the challenge of expensive and non-trivial structure-property mapping in organic photovoltaics (OPVs), where morphology strongly influences power conversion efficiency.
• Overcome limitations of traditional methods—such as PDE simulations and intuitive descriptors—that are either computationally prohibitive or insufficiently comprehensive.
• Develop a data-driven, interpretable deep learning framework (DLSP) to accelerate and guide the discovery of high-performance OPV microstructures.
• Enable both manual and automated design of optimal morphologies by leveraging a trustworthy, fast surrogate model.
• Demonstrate that interpretability of the model can reveal physically meaningful features, such as interfacial connectivity and domain size, that correlate with performance.

Proposed method

• Train a custom, shallow CNN to classify OPV microstructures into 10 performance bins based on short-circuit current density ($J_{sc}$), using morphologies generated via kinetic Monte Carlo simulations.
• Validate the model using in-sample and out-of-sample datasets to ensure robustness and generalization, with no severe overfitting observed.
• Use saliency maps to interpret the model’s predictions and identify the most influential morphological features, such as interfacial area and domain connectivity.
• Integrate the trained CNN as a fast, differentiable cost function in a population-based incremental learning (PBIL) optimization framework for automated morphology design.
• Employ PBIL to evolve a probabilistic morphology representation (a 2D probability matrix $P$) by iteratively updating it based on the performance of sampled morphologies, reinforcing high-performing features.
• Use embarrassingly parallel evaluation of multiple morphologies to scale the optimization process efficiently.

Experimental results

Research questions

• RQ1 Can a deep learning model serve as an accurate and generalizable surrogate for the complex, non-linear structure-property relationship in OPV active layer morphologies?
• RQ2 How do interpretability techniques like saliency maps reveal physically meaningful features in OPV microstructures that are not captured by standard pre-trained networks?
• RQ3 Can a fast, interpretable deep learning model enable effective manual and automated design of high-performance OPV morphologies?
• RQ4 To what extent does a custom-designed CNN outperform standard architectures like VGG-16 and ResNet-50 in both accuracy and feature interpretation for this specific materials science problem?
• RQ5 How can physics-informed priors be integrated into the training process to reduce data requirements and improve model robustness?

Key findings

• The custom CNN achieved high accuracy in classifying $J_{sc}$ performance levels and generalized well to out-of-sample morphologies, with no significant overfitting observed.
• Saliency maps revealed that the custom CNN consistently identified critical interfacial features—such as connected pathways and domain boundaries—as key drivers of performance, whereas VGG-16 and ResNet-50 failed to do so reliably.
• The PBIL-optimized morphologies evolved toward fine, finger-like fractal structures resembling those reported in prior studies, indicating effective exploration of the performance manifold.
• The custom CNN outperformed VGG-16 and ResNet-50 in both classification accuracy and interpretability, despite being shallower, demonstrating the value of task-specific architecture design.
• The integration of the trained model into the PBIL framework enabled efficient, parallelized optimization of morphologies, accelerating the discovery of high-performance structures.
• The framework successfully demonstrated that interpretability and speed can coexist in deep learning models for materials science, enabling both insight and design.