PhysSAE: Mechanistic Interpretability of PINNs with Sparse Autoencoders
Abstract
Physics-Informed Neural Networks (PINNs) embed PDE residuals into neural network training, but their internal representations remain opaque: it is unknown what physical features their hidden layers encode or whether those features have a localized causal role. We present PhysSAE, a mechanistic interpretability framework that trains overcomplete sparse autoencoders (SAEs) on PINN penultimate-layer activations and evaluates dictionary atoms through direct causal intervention in the original frozen hidden state: hcf = h − αzk dk, bypassing the SAE decoder entirely. Across six PDE families, with 3 PINN seeds and 3 SAE seeds each---we show that (i) Our discovered SAE atoms align with independently-defined physical observables (max Pearson |r|=0.951, always ≫ permutation null), (ii) the causal footprint of top-aligned atom ablation is 1.2--4.2× more spatially concentrated canonical than PCA or ICA interventions, and (iii) top-aligned atoms outperform matched random controls on causal localization for structured physical concepts (ESF80 advantage 0.04-0.44). Two-atom bilateral representations improve concept regression R2 by ΔR2 = 0.05−0.15 over single atoms, while random pairs decrease it by up to 0.60. These results demonstrate that PINNs develop sparse, physically structured latent representations that can be identified and causally interrogated post-hoc, opening a path toward interpretability-aware scientific machine learning.
Related
- Flag Game: A Toy Model for Mechanistic Swarm Interpretability
- Sparse data limit what a mechanistic corrosion model can predict
- Derivative Computation in PINNs: Automatic Differentiation, Finite Differences and Beyond
- Accelerating natural gradient descent for PINNs with randomized numerical linear algebra
- An improved periodic activation for PINNs reconstructing convective flows
BibTeX
@misc{patil2026physsae,
title = {{PhysSAE: Mechanistic Interpretability of PINNs with Sparse Autoencoders}},
author = {Nandita N. Patil and Eshwar R. A. and Gajanan V. Honnavar},
howpublished = {arXiv},
year = {2026},
month = sep,
doi = {10.48550/arxiv.2609.07061},
eprint = {2609.07061},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2609.07061},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - PhysSAE: Mechanistic Interpretability of PINNs with Sparse Autoencoders
AU - Nandita N. Patil
AU - Eshwar R. A.
AU - Gajanan V. Honnavar
PY - 2026
DA - 2026/09/12
PB - arXiv
DO - 10.48550/arxiv.2609.07061
UR - https://arxiv.org/abs/2609.07061
AB - Physics-Informed Neural Networks (PINNs) embed PDE residuals into neural network training, but their internal representations remain opaque: it is unknown what physical features their hidden layers encode or whether those features have a localized causal role. We present PhysSAE, a mechanistic interpretability framework that trains overcomplete sparse autoencoders (SAEs) on PINN penultimate-layer activations and evaluates dictionary atoms through direct causal intervention in the original frozen hidden state: h_cf = h − αz_k d_k, bypassing the SAE decoder entirely. Across six PDE families, with 3 PINN seeds and 3 SAE seeds each---we show that (i) Our discovered SAE atoms align with independently-defined physical observables (max Pearson |r|=0.951, always ≫ permutation null), (ii) the causal footprint of top-aligned atom ablation is 1.2--4.2× more spatially concentrated canonical than PCA or ICA interventions, and (iii) top-aligned atoms outperform matched random controls on causal localization for structured physical concepts (ESF_80 advantage 0.04-0.44). Two-atom bilateral representations improve concept regression R^2 by ΔR^2 = 0.05−0.15 over single atoms, while random pairs decrease it by up to 0.60. These results demonstrate that PINNs develop sparse, physically structured latent representations that can be identified and causally interrogated post-hoc, opening a path toward interpretability-aware scientific machine learning.
ER -