Physics-informed time-series forecasting of perovskite photoluminescence stability
Physics-informed time-series forecasting of perovskite photoluminescence stability
Authors: Alexander Wieczorek, Manuel Kober‐Czerny, Fábio Lopes, Austin George Kuba, L. Müller, Christian M. Wolff, Jason Hattrick-Simpers, Sebastian Siol Date: 2026-06-11 Paper ID: openalex:2606.13414
Summary
This paper presents a physics-informed forecasting framework to predict the long-term photoluminescence (PL) stability of metal halide perovskites under accelerated aging conditions. The method featurizes PL spectra using domain-specific physical models and employs a hybrid CNN-LSTM architecture for time-series forecasting. Evaluation on a dataset of 167 perovskites shows that the model achieves superior performance over baselines and provides interpretable stability descriptors, facilitating faster material discovery and testing.
Key Contributions
- Introduces a physics-informed time-series forecasting framework for predicting long-term photoluminescence stability in metal halide perovskites.
- Curates a dataset of 167 perovskite materials comprising over 86k photoluminescence spectra under heat and light stress.
- Demonstrates that combining physics-based featurization with a hybrid CNN-LSTM architecture outperforms baseline benchmarks while maintaining interpretability.
Key Concepts
- physics-informed-time-series-forecasting: A forecasting framework that integrates physics-based featurization with neural architectures to predict temporal dynamics in material science.
Archivist Review
The core contribution of physics-informed feature engineering for time-series forecasting is a valuable, reusable methodology. The hybrid CNN-LSTM architecture was rejected as it represents a generic implementation detail rather than a distinct, novel framework concept. No datasets or open questions were sufficiently distinct or critical to warrant permanent vault inclusion.
Approved Concepts
- Physics-informed time-series forecasting: The core methodological contribution combines domain-specific physics knowledge with forecasting models to improve predictive performance and interpretability in material science applications.
Rejected Candidates
- [concept] Hybrid CNN-LSTM architecture (
hybrid-cnn-lstm-architecture) - not_novel: This is a routine neural network design rather than a novel or highly specific forecasting methodology.
Links
Metadata & Links
- url
- https://arxiv.org/abs/2606.13414
- paper_id
- 2606.13414
- paper_source
- openalex
- domain
- nlp
- tags
-
- architectures
-
- datasets
-
- concept_slugs
- physics-informed-time-series-forecasting
- dataset_slugs
-
- skill
- TimeSeriesSkill
- created_at
- 2026-06-14T06:33:08Z