Making Neural Networks Physically Interpretable

674 concepts590 papers5 proposals

Deep neural networks excel at predicting complex physical phenomena, from fluid dynamics to battery degradation. However, in scientific domains, a model that merely predicts an outcome without explaining why is often insufficient. As noted in Cycle Life Prediction for Lithium-ion Batteries: Machine Learning and More, purely data-driven models operate as black boxes, failing to reveal the underlying physical mechanisms. This lack of transparency creates a "physical interpretability gap," hindering scientific validation and hypothesis generation (Deep learning in fluid dynamics; MRS Communications Abstracts).

To bridge this gap, a growing body of research focuses on designing neural network architectures and training paradigms that explicitly surface human-understandable physical principles, such as conservation laws, symmetries, and differential equations.

Structuring Networks with Mathematical Inductive Biases

Instead of treating neural networks as arbitrary function approximators, researchers embed classical mathematical structures directly into the network's architecture. This forces the model's internal representations to align with physical reality.

For instance, in PDE-Net: Learning PDEs from Data, Yiping Lu et al. represent differential operators as constrained convolution kernels. By doing so, the network does not just fit data; it learns the explicit partial differential equations (PDEs) governing the system. Similarly, in TaylorNet: A Taylor-Driven Generic Neural Architecture, Hongjue Zhao et al. replace standard non-linear activation functions with Taylor polynomials, making the learned representations mathematically traceable. For uncovering fundamental system properties, Nicolas Boullé et al. utilize rational neural networks in Data-driven discovery of Green’s functions with human-understandable deep learning to discover Green's functions, revealing singular behaviors and conservation laws directly from data.

Discovering Symmetries and Conserved Quantities

Physical systems are governed by invariants—quantities like energy or momentum that remain constant over time. To extract these from black-box models, researchers use specialized network topologies.

In Discovering Symmetry Invariants and Conserved Quantities by Interpreting Siamese Neural Networks, Sebastian Johann Wetzel et al. demonstrate that Siamese Neural Networks equipped with a strategic "bottleneck" layer can automatically identify symmetry invariants and conserved quantities. By training the network to determine whether two distinct states belong to the same physical trajectory, the bottleneck layer is forced to compress the data down to its fundamental constants, such as total energy or spacetime intervals.

Distilling Neural Weights into Symbolic Formulas

Another powerful approach involves translating the continuous, uninterpretable weights of a trained neural network into discrete, human-readable mathematical formulas.

In PySINDy: A Python package for the sparse identification of nonlinear dynamical systems from data, Brian M. de Silva et al. present a framework that frames model discovery as a sparse regression problem over a library of candidate physical terms, ensuring the resulting equations adhere to fundamental conservation laws.

To tackle high-energy physics data, Anja Butter et al. apply symbolic regression in Back to the formula - LHC edition to extract explicit kinematic formulas from complex Large Hadron Collider simulations, turning an opaque neural predictor into an analytical expression. Similarly, in Discovering interpretable elastoplasticity models via the neural polynomial method enabled symbolic regressions, Bahador Bahmani et al. combine neural polynomials with symbolic regression to discover interpretable material models, allowing engineers to analytically validate safety properties like convexity.

Orthogonal Regularization and Hybrid Modeling

When combining deep learning with known first-principles physics, a common failure mode is that the neural network redundantly "re-learns" the physics already captured by classical equations, obscuring its unique contributions.

To resolve this, Maarten Schoukens et al. introduce an orthogonal projection-based regularization technique in Orthogonal projection-based regularization for efficient model augmentation. This method penalizes the deep learning component if it attempts to model dynamics already captured by the first-principles model. By forcing the neural network to operate exclusively in the orthogonal complement of the known physics, the hybrid model remains highly interpretable, isolating exactly what the classical model missed.

Why It Works: Restricting the Hypothesis Space

Why do these methods successfully bridge the physical interpretability gap?

Standard neural networks have an immense hypothesis space, allowing them to find millions of mathematically valid but physically nonsensical ways to fit a dataset. By introducing physical constraints—whether through constrained convolution kernels, sparse symbolic libraries, or architectural bottlenecks—the hypothesis space is drastically restricted to manifolds that conform to physical laws.

When the network is forced to solve a problem using only the building blocks of mathematical physics (such as derivatives, polynomials, or invariants), its internal parameters lose their opacity. The weights cease to be arbitrary statistical coefficients and instead map directly onto physical properties, enabling scientists to validate, trust, and derive new theoretical insights from machine learning models.

History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_model_interpretability_black_box_model_o_b594f97b,
  title = {model interpretability (black-box model opacity)},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/model-interpretability-black-box-model-opacity/rdr_b594f97b75db4cb6be87d11a90b1cd35},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_model_interpretability_black_box_model_o_b594f97b_v1,
  title = {model interpretability (black-box model opacity)},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/model-interpretability-black-box-model-opacity/rdr_b594f97b75db4cb6be87d11a90b1cd35/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake