Research

Reliable learning for physical dynamics and control.

I study when neural surrogates preserve the state evolution, local sensitivities, and closed-loop behavior required for model-based control.

A central question is when a learned dynamics model is reliable enough for controller design.

Low state-prediction error is not sufficient for control. A useful surrogate must also preserve action-conditioned dynamics and the local sensitivities used by an optimizer or policy.

I study these requirements through neural operators, model-based reinforcement learning, direct numerical simulation, and controlled perturbation tests.

In preparation for submission to TMLR

Audited Surrogate Gradients for Efficient Model-Based Flow Control

With Stefania Fresca
University of Washington

Workshop version: Accepted to the Sim2Science workshop at NeurIPS 2026. Journal manuscript in preparation for Transactions on Machine Learning Research (TMLR).

The question

Accurate state prediction does not imply a faithful control Jacobian.

Neural surrogates can reduce the cost of trajectory optimization, but controller design depends on derivatives with respect to the action. These derivatives can be wrong even when one-step state errors are small.

I compare automatic-differentiation gradients from the surrogate with central finite-difference sensitivities from a high-fidelity solver on held-out states. A compact feedback law fitted to the measured gradient field initializes a neural policy, and an anchor penalty keeps optimization close to that law. Controllers are evaluated in the original solver across four test cases.

Physical flow fields, surrogate predictions, and errors shown at the initial state, training boundary, and a later diagnostic time — Teacher-forced predictions remain accurate near the training horizon, while the autoregressive rollout later diverges from DNS. The distinction bounds which dynamical claims the surrogate can support.

Perturb

Estimate action sensitivities with paired finite-difference DNS trajectories.

Audit

Compare gradient sign, direction, and magnitude across relevant horizons.

Distill

Fit a compact feedback law to the measured high-fidelity gradient field.

Refine

Initialize a neural policy from the feedback law and keep it close with an anchor penalty.

Validate

Evaluate closed-loop performance and falsification controls in DNS.

Physical test

The audited gradient identifies a useful feedback phase.

The local action gradient predicts a feedback phase that lowers the objective in direct numerical simulation. A falsification phase produces the opposite response, providing a controlled test of the inferred mechanism.

A phase sweep and closed-loop physical simulation comparing a controller extracted from audited gradients with a falsification phase — The phase selected from the audited gradient lowers the physical objective over the evaluation horizon.

Research principle

Prediction accuracy, action-gradient fidelity, and closed-loop performance are distinct properties. An audit is local to the sampled states and actions, so gradients must also be measured around the actions the deployed controller applies.

Submitted to L4DC · Learning for Dynamics and Control Conference (L4DC)

Spatial Context or Memory? Sizing the Inputs of Neural Surrogates for Partially Observed Fields

When can a longer sensor history compensate for a smaller spatial observation window? I study this trade-off through exact Gaussian conditioning and learned local models on a forced linear field, viscous Burgers flow, and the chaotic Kuramoto–Sivashinsky equation.

Noisy or sparse observations alone do not force a predictor to use distant sensors: spatial correlation in the underlying field matters. How much history can replace spatial context depends on the window size, sensor spacing, transport, and forecast lead time. In the tested Kuramoto–Sivashinsky regimes, history becomes more useful at longer lead times for windows up to one correlation length, a pattern absent from the tested Burgers regimes.

arXiv version

Partial observations
Temporal memory
Spatial context
PDE forecasting

Environmental prediction

Interpretable deep learning for wildfire spread

We trained Autoencoder, ResNet, U-Net, and Swin-UNet architectures for next day fire-mask segmentation using nearly a decade of remote sensing data from California.

We evaluated predictive skill under class imbalance, computational cost, and data efficiency. SHAP, Grad-CAM, and Integrated Gradients were used to compare global and spatial feature attribution.

Read the paper Code and data

Five remote sensing samples with twelve input fields, a previous fire mask, and the next day fire mask — Each row is one sample. The inputs include weather, terrain, vegetation, population density, energy release, and the previous fire mask. The final column is the next day fire mask. Figure from the published paper.

SHAP feature importance rankings for XGBoost, Autoencoder, ResNet, U-Net, and Swin-UNet wildfire models — Mean absolute SHAP values show that each architecture relies on a different mix of environmental inputs. Figure from the published paper.

Main results

U-Net and Swin-UNet achieved the highest AUC-PR.

Swin-UNet reached a mean AUC-PR of 0.2803 and U-Net reached 0.2739. Both outperformed the Autoencoder and ResNet baselines on the highly imbalanced segmentation task.

The attribution profiles differed despite similar aggregate performance. U-Net emphasized population density and the previous fire mask, while Swin-UNet emphasized vegetation, humidity, and the previous fire mask.

The result shows why predictive metrics and attribution analysis should be considered together when comparing architectures.

Additional research

Machine learning for satellite dynamics

I also contribute to the FISICA project on machine learning for satellite dynamics. Further details are limited due to confidentiality.