Short notes on methodology, deployment, and research direction. Updated occasionally.
May 2026 · Methodology
The trade-off between physical plausibility and raw accuracy, and why adaptive penalty design matters.
A common pattern in physics-informed learning is to add a physics penalty term directly to the training loss. The model minimizes a weighted sum of data error and violation of some physical constraint — conservation of mass, momentum balance, an energy budget. In principle this is elegant: the model learns from data but is nudged toward physically plausible behavior.
In practice this often hurts raw predictive accuracy. When we force a data-driven forecaster to satisfy an approximate or simplified equation, we inject bias. The model starts optimizing to satisfy the hand-written physics term instead of strictly minimizing forecast error. This is especially visible in the atmospheric boundary layer and dense urban canopies, where simplified equations don't capture the real physics.
The frontier is adaptive loss functions that apply physics pressure selectively — in high-risk regimes (gust fronts, nocturnal heat traps) where getting the physics wrong has operational consequences, while relaxing the penalty in calm regimes where data alone is sufficient. The key insight: the penalty design itself becomes part of the model architecture and must co-evolve with deployment constraints.
A full CFD solve is not compatible with a 200 ms inference budget on a substation controller with sub-2 GB memory. We need a deployable physics prior — not an oversimplified analytic penalty and not a full-blown simulator.
April 2026 · Deployment
Lessons from deploying wind forecasting models on edge hardware with strict latency and memory budgets.
Deploying wind forecasting models at the turbine edge is qualitatively different from running them on a workstation. The constraints are real: sub-200 ms inference latency, sub-2 GB system RAM, and power budgets that rule out GPU acceleration. Here's what I learned from the shallow hybrid model work.
We found that inference latency, not model size, is the binding constraint. A compact LSTM with 50K parameters can still exceed 200 ms if the input preprocessing is heavy or if memory access patterns are poor. The bottleneck is often in the data pipeline — feature normalization, sliding window assembly, and tensor reshaping.
A model that produces a forecast 95% of the time with 2 m/s RMSE is more valuable than one that produces a forecast 70% of the time with 1.5 m/s RMSE. Edge deployments need graceful degradation — if a sensor drops or latency spikes, the model should still produce a usable output, not crash.
March 2026 · Direction
Why the global research funding vocabulary is shifting, and what it means for physics-informed AI.
Sustainable Development Goal (SDG) language still matters — it provides a public-good framework that connects research to societal outcomes. But funding agencies and national programs are increasingly asking whether a method improves resilience, security, and infrastructure capacity rather than whether it supports a specific SDG target.
For physics-informed AI researchers, this shift is useful rather than threatening. The same model that supports clean energy (SDG 7) can also support grid reliability. The same heat-risk forecast (SDG 11) can also support emergency response and city operations. The practical target becomes an operational model that remains physically credible when measurements are noisy, regimes change, and decisions must be made locally.
The strongest future framing is not "SDGs versus security." It is SDGs translated into resilience, energy independence, and infrastructure intelligence.