⚡ electrical engineering

PHAST: Port-Hamiltonian Architecture for Structured Temporal Dynamics Forecasting

PHAST is a port-Hamiltonian deep learning architecture that leverages Strang splitting and structured Hamiltonian decomposition to achieve stable long-horizon forecasting and physically meaningful parameter recovery from position-only observations across diverse dissipative physical systems.

Original authors: Shubham Bhardwaj, Chandrajit Bajaj

Published 2026-02-23

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Shubham Bhardwaj, Chandrajit Bajaj

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to predict the future path of a swinging pendulum, a bouncing ball, or even the population of rabbits and foxes in a forest. You can see where they are (their position), but you can't see how fast they are moving or how much energy they have left. This is the challenge of forecasting in the real world: we only get partial glimpses of reality, yet we need to predict what happens next.

Most current AI models try to learn these patterns by memorizing data, like a student cramming for a test. They often work well for a short time but eventually "hallucinate" or go crazy when asked to predict far into the future. They forget the basic laws of physics, like how friction slows things down or how energy is conserved.

Enter PHAST (Port-Hamiltonian Architecture for Structured Temporal dynamics). Think of PHAST not as a student who memorizes, but as a physics-savvy engineer who understands the rules of the game before looking at the data.

Here is how PHAST works, broken down into simple concepts:

1. The "Energy Bank" Analogy

Imagine every physical system (a pendulum, a circuit, a robot arm) has an Energy Bank.

Conservative Systems (The Savings Account): In a perfect world with no friction, energy just moves back and forth. It's like moving money between a checking and savings account; the total amount never changes.
Dissipative Systems (The Leaky Bucket): In the real world, systems lose energy. A pendulum slows down because of air resistance; a battery loses charge as heat. This is like a bucket with a hole in it. The water (energy) level must go down over time.

Most AI models treat the bucket as if it's sealed. PHAST, however, is built with a leaky bucket in mind. It knows that energy must decrease or stay the same, never magically increase. This "leak" is built into its brain, ensuring it never predicts a pendulum swinging higher and higher forever.

2. The Three "Knowledge Regimes"

PHAST is flexible. It can work whether you give it a full physics textbook, just a few hints, or no hints at all. The authors call these three modes:

The "Known" Mode (The Physics Expert): You tell PHAST exactly how gravity works and how heavy the object is. It only has to learn the "leak" (friction). It's like giving a mechanic the blueprints and asking them to find the broken part.
The "Partial" Mode (The Detective): You give PHAST the shape of the potential energy (like "gravity pulls down") but not the exact numbers. It has to figure out the mass and the friction. It's like a detective who knows the crime scene layout but has to find the suspect.
The "Unknown" Mode (The Black Box): You give PHAST nothing but a video of the movement. It has to invent the physics from scratch. Even here, PHAST forces itself to obey the rule that "energy cannot be created," which keeps it from going crazy.

3. The "Householder" Trick (The Efficient Toolbox)

To make sure the math works correctly, PHAST uses a clever trick called Householder-style parameterization.

The Problem: If you ask a standard AI to learn "friction," it might accidentally learn "anti-friction" (which would make things speed up forever), breaking the laws of physics.
The Solution: PHAST builds its friction component like a stack of positive weights. It's mathematically impossible for it to create negative friction. It's like building a wall out of bricks that only stack up; you can't accidentally build a hole that goes down. This guarantees the system stays stable.

4. The "Splitting" Strategy (The Dance)

To predict the future, PHAST doesn't try to solve a giant, messy equation all at once. Instead, it uses a technique called Strang Splitting.

Imagine a dancer moving in a circle (conservative motion) while slowly losing energy (dissipation).
Instead of trying to do both at once, PHAST takes a tiny half-step of "losing energy," then a full step of "dancing," then another half-step of "losing energy."
This simple dance step-by-step approach keeps the prediction accurate and stable for a long time, whereas other models might trip over their own feet after a few steps.

5. Why This Matters: The "Two-Axis" Test

The paper makes a brilliant observation: Being good at predicting the path is different from being good at understanding the physics.

A model could cheat: It might predict the pendulum's path perfectly by faking the friction numbers to cancel out its own mistakes. It looks good on the scoreboard (forecasting) but is physically wrong (identifiability).
PHAST separates these two goals. It checks: "Did you predict the path?" AND "Did you learn the actual friction?"
In their tests, PHAST didn't just predict better than other AI models; it actually recovered the true physical parameters (like the exact amount of air resistance) when given enough hints.

The Big Picture

PHAST is a new kind of AI that respects the laws of physics by design. It doesn't just guess what happens next; it simulates the underlying energy flow. Whether you are modeling a robot arm, an electrical circuit, or a predator-prey ecosystem, PHAST ensures that the future it predicts is stable, realistic, and physically possible.

In short: PHAST is the AI that knows the rules of the universe before it starts playing the game.

1. Problem Statement

The paper addresses the challenge of forecasting the dynamics of dissipative physical systems from position-only observations (the $q$ -only setting).

The Challenge: Real-world systems (pendulums, circuits, ecological models) lose energy over time due to friction, resistance, or heat. Standard physics-informed neural networks (like HNNs) often assume conservative systems (energy-preserving), failing to model dissipation accurately. Furthermore, learning dynamics from positions alone ( $q_t$ ) without observing momenta ( $p_t$ ) creates an ill-posed inverse problem. Multiple combinations of mass, potential energy, and damping can produce identical trajectories (gauge freedom), making physical parameter recovery difficult even if forecasting is accurate.
The Goal: Develop a model that (a) produces stable, long-horizon open-loop forecasts and (b) recovers physically meaningful parameters (potential $V$ , mass $M$ , damping $D$ ) when partial physical knowledge is available.

2. Methodology: PHAST Architecture

PHAST (Port-Hamiltonian Architecture for Structured Temporal dynamics) is built on the Port-Hamiltonian (pH) framework, which explicitly decomposes system dynamics into conservative and dissipative components.

A. Mathematical Foundation

The system dynamics are modeled as:
$\dot{x} = (J - R)\nabla H(x)$
Where:

$x = (q, p)$ is the phase state (position and momentum).
$H(x) = V(q) + \frac{1}{2}p^\top M(q)^{-1}p$ is the Hamiltonian (Total Energy), composed of potential energy $V$ and kinetic energy defined by mass/inertia $M$ .
$J$ is the skew-symmetric interconnection matrix (conservative energy exchange).
$R = \text{diag}(0, D(q))$ is the symmetric positive semi-definite (PSD) dissipation matrix, where $D(q)$ acts only on momentum (damping).
Passivity Guarantee: By construction, $\frac{dH}{dt} = -\nabla H^\top R \nabla H \leq 0$ , ensuring energy non-increase and system stability.

B. Three Knowledge Regimes

PHAST unifies learning across three regimes by swapping how the components $(V, M, D)$ are instantiated:

KNOWN: Physics is fully known ( $V$ and $M$ are fixed functions); only damping $D(q)$ is learned.
PARTIAL: The functional form of $V$ or $M$ is known (e.g., from CAD or first principles), but parameters are learned. $D(q)$ is learned with bounded strength.
UNKNOWN: All components are learned via neural networks, constrained only by the pH structure (SPD for $M$ , PSD for $D$ ).

C. Key Technical Components

$q$ -Only Pipeline: Since momenta are latent, PHAST uses a causal velocity observer (Finite Differences + Temporal Convolutional Network) to estimate $\dot{q}$ from position history. A canonicalizer maps this to a phase state $(q, \hat{p})$ .
Householder-Style Parameterizations: To ensure $M(q) \succ 0$ $M (q) ≻ 0$ and $D(q) \succeq 0$ $D (q) ⪰ 0$ efficiently, PHAST uses low-rank outer-product expansions:
- $D(q) = d_0 I + \sum \beta_i k_i k_i^\top$ (PSD).
- $M(q) = \Lambda + UU^\top$ (SPD).
- This allows efficient matrix-vector products ($O(nr)$) and Woodbury identity for inversion ( $O(nr^2)$ ).
Damping Bounds: To prevent the learned damping $D$ from acting as an "error sink" (absorbing model mismatch rather than representing true physics), PHAST optionally bounds the total damping strength ( $\sum \beta_i \leq \bar{\beta}$ ). This is crucial for identifiability.
Structure-Preserving Integration: The model uses Strang Splitting (Leapfrog for conservative part + half-step damping) to advance dynamics. This preserves the continuous-time dissipation structure to $O(\Delta t^2)$ .

3. Key Contributions

Unified Architecture: A single pH template handling KNOWN, PARTIAL, and UNKNOWN regimes, bridging the gap between pure physics and black-box learning.
Efficient Structured Primitives: Novel low-rank parameterizations that guarantee positive definiteness of mass and damping matrices while maintaining computational efficiency.
Identifiability via Bounding: Demonstrates that unconstrained dissipation leads to non-identifiable parameters. Introducing damping bounds breaks gauge freedom, enabling the recovery of physically meaningful damping fields.
Two-Axis Evaluation: Proposes separating forecasting stability (Rollout MSE) from physical identifiability (Damping $R^2$ ). The paper shows that high forecasting accuracy does not guarantee correct parameter recovery without structural anchors.
Control Capability: Extends the framework to Energy-Casimir control, demonstrating that the learned pH structure can be used for passivity-based stabilization in closed-loop settings, provided the port output (velocity) is estimated accurately.

4. Experimental Results

The authors evaluated PHAST on 13 benchmarks across six domains: Mechanical (pendulums, cart-pole, double pendulum), Electrical (RLC), Molecular (Lennard-Jones), Thermal (heat exchange), Astrophysical (N-body), and Ecological (predator-prey).

Forecasting Stability: PHAST significantly outperforms baselines (S5, LinOSS, Transformers, VPT) in long-horizon open-loop rollouts.
- Example: On the Windy Pendulum, PHAST (KNOWN) reduced rollout MSE by 26.5x compared to the best baseline (D-LinOSS).
- PHAST maintains stability in chaotic systems (Double Pendulum) where baselines diverge quickly.
Parameter Recovery:
- In KNOWN regimes, PHAST recovers the true position-dependent damping with $R^2 \approx 0.996$ .
- In PARTIAL regimes, applying damping bounds allows for meaningful recovery of friction parameters, whereas unbounded models fail to identify them ( $R^2 \ll 0$ ) despite accurate rollouts.
- In UNKNOWN regimes, forecasting remains stable, but parameter recovery is inherently ill-posed without anchors.
Control: In closed-loop stabilization tasks, PHAST-based controllers achieved 100% success rates. The study highlighted that the bottleneck in $q$ -only control is the quality of the velocity estimate (port output), not the model structure itself.

5. Significance and Impact

Scientific Machine Learning: PHAST provides a rigorous framework for learning dissipative systems, moving beyond the conservative assumptions of HNNs and LNNs.
Interpretability: By enforcing physical constraints (passivity, positive definiteness) and using structured parameterizations, PHAST enables the recovery of interpretable physical parameters (mass, damping) rather than just "black box" predictions.
Robustness: The architecture guarantees stability (non-increasing energy) by construction, making it suitable for safety-critical applications like robotics and control.
Generalizability: The framework is domain-agnostic, successfully applied to mechanical, electrical, thermal, and ecological systems, proving the universality of the Port-Hamiltonian decomposition.

In summary, PHAST demonstrates that structural priors (Port-Hamiltonian decomposition) are more critical than model capacity for long-horizon stability and physical interpretability in dissipative systems, provided that appropriate constraints (like damping bounds) are applied to resolve identifiability issues.