Abstract
Energy conservation is a foundational principle in physics, yet real-world systems often exhibit non-conservative behavior due to friction, external bodies, or radiation. We present the Neural New-Physics Detector (NNPhD), a machine learning framework that identifies "new physics" by decomposing unknown force fields into conservative and non-conservative components. By exploiting a universal phase transition in the model's loss function, NNPhD can automatically detect when energy conservation is violated. We demonstrate its efficacy by rediscovering historical breakthroughs—including friction, the existence of Neptune, and gravitational waves—directly from raw trajectory data.
Introduction
In the history of physics, apparent violations of energy conservation have frequently led to major discoveries. When a system's energy doesn't add up, physicists search for "new physics"—be it unseen particles like neutrinos, hidden celestial bodies like Neptune, or dissipative forces like friction.
Current physics-informed machine learning models, such as Lagrangian Neural Networks (LNNs), are designed to strictly conserve energy. While powerful, they fail when applied to dissipative or externally influenced systems. Conversely, "black-box" neural networks can fit any data but lack the interpretability and generalization required for scientific discovery. NNPhD addresses this gap, providing a method to automatically detect and explain non-conservative components without requiring the exact form of the "old" conservative physics to be known beforehand.
Methodology: The NNPhD Architecture
The Neural New-Physics Detector (NNPhD) utilizes a dual-module architecture to model complex dynamics. As shown in the figure below, the system consists of two parallel neural networks:
- Lagrangian Neural Network (LNN): This module parametrizes the "old" or conservative physics. It outputs a force field derived from a learned Lagrangian, ensuring the predicted dynamics obey energy conservation by construction.
- Universal Approximator Network (UAN): A standard multi-layer perceptron (MLP) that captures the residual, non-conservative forces—the "new physics."
The model is trained to minimize the total force recovery error while applying a regularization term $\lambda$ to the magnitude of the non-conservative force. This setup allows the model to "decide" how much of the observed dynamics should be attributed to known conservative laws versus unknown external or dissipative factors.
The Universal Phase Transition
The core innovation of NNPhD is the discovery of a universal "phase transition" in the model's behavior. The authors prove that a sharp transition occurs at a regularization weight of $\lambda = 1$.
- When $\lambda < 1$, the model correctly decomposes the force into its conservative and non-conservative parts.
- When $\lambda > 1$, the regularization is too strong, forcing the model to attempt to fit the entire dynamics into the conservative LNN module.
As illustrated in the plot, this jump in prediction error at $\lambda=1$ acts as a diagnostic tool. If a researcher scans $\lambda$ and observes this sharp increase in error, it provides definitive evidence that the system contains non-conservative "new physics."
Experimental Results and Historical Rediscovery
The NNPhD framework was evaluated on several simulated systems representing landmark historical discoveries:
- Damped Double Pendulum: Rediscovering friction (first recorded by Da Vinci).
- Uranus’ Orbit: Rediscovering Neptune by treating its gravitational pull as a non-conservative "new" force relative to a Sun-Uranus model.
- Binary Black Hole Inspiral: Rediscovering gravitational radiation by modeling energy loss as a dissipative force.
Key Insights
- Discovery without Prior Knowledge: Unlike historical physicists who needed a specific functional form for gravity to find residuals, NNPhD learns the conservative part from scratch.
- Interpretability: By passing the UAN's output through a symbolic regression module, the authors converted black-box weights into human-readable formulas (e.g., $f = -\gamma \dot{q}$), successfully extracting correct symbolic expressions for friction and orbital parameters.
- Superior Extrapolation: In tasks like the damped pendulum, NNPhD significantly outperformed pure LNNs (which cannot model decay) and black-box models (which fail to generalize). It correctly predicted exponential energy decay even on unseen test data.
Limitations and Future Work
While promising, NNPhD requires high-quality data coverage. If data is limited—for instance, observing a damped oscillator moving in only one direction—the damping may be misidentified as a conservative constant force. Additionally, the symbolic "explainer" stage currently benefits from human-provided feature sets. Future research will focus on scaling this approach to higher-dimensional systems, such as complex fluid dynamics or high-energy particle physics.
Summary updated