Imagine two students learning to draw human figures. The first memorizes ten thousand photographs and reproduces them convincingly, right up until you ask for a pose that wasn’t in the album, at which point elbows start bending backward. The second learns skeletal anatomy first, and every drawing they make afterward has joints that work, because the bones were never optional. Most deep learning is the first student. Scientific machine learning is a long argument for the second.
That argument got a visible endorsement this year: an international team including Prof. Dr. Markus Lange-Hegermann had its paper selected as a Spotlight at NeurIPS 2026. The paper introduces FLASH-MAX, a machine learning architecture that combines neural networks with Maxwell’s equations and reconstructs electromagnetic fields accurately. That’s the extent of what’s been made public so far, and I want to be careful not to pad it with invented internals. But the shape of the result is what interests me, and the shape is worth talking about on a site concerned with agent architecture.
Two ways to make a model behave
There are broadly two places you can put a constraint in a learning system. You can put it in the loss, where it becomes a preference the optimizer trades against everything else. Or you can put it in the architecture, where it becomes a property of the function class itself — something the model cannot violate because violations aren’t representable.
Physics-informed neural networks, the best-known family in this space, mostly live in the first category. You add residual terms for the governing equations and let gradient descent negotiate. It works, and it has produced real science, but anyone who has trained a PINN knows the negotiation can go badly: stiff loss surfaces, competing terms, and training runs that converge to something plausible-looking and physically wrong. A good deal of recent work is about making that training tractable rather than heroic. Ben Moseley’s ELM-FBPINNs, for instance, significantly accelerated PINN training by combining the networks with multiple levels of domain decomposition — a structural intervention aimed squarely at an optimization problem.
When a paper is framed as combining neural networks with Maxwell’s equations and gets highlighted for accurate field reconstruction, the interesting question is which category it falls into. Accuracy on field reconstruction is a demanding target. Electromagnetic fields are not forgiving: divergence conditions and curl relationships couple everything to everything, and small inconsistencies propagate into physically absurd predictions. A model that reconstructs them well is a model whose errors have nowhere convenient to hide.
Why agent people should care
Agent systems are in the loss-term phase of their development. We express nearly every desirable property as a soft preference: a system prompt, a reward signal, a post-hoc validator, a retry loop. The agent is asked to prefer valid tool calls, prefer honest citations, prefer staying inside its authorized scope. When it fails, we add another preference.
SciML has been running the alternative experiment for years, and its lesson is fairly blunt. Constraints encoded in structure are cheap at inference and free at evaluation. Constraints encoded in objectives cost you training stability and buy you no guarantees. The reason a field-reconstruction result lands at a top venue is not that the network is large. It’s that someone worked out how to carry known structure into the model rather than reminding the model of that structure ten thousand times and hoping.
Translating that into agent design looks something like:
- Typed, schema-constrained action spaces instead of prompts asking for well-formed calls
- Capability scoping enforced at the tool boundary instead of instructions not to exceed scope
- State machines that make illegal transitions unreachable instead of validators that catch them afterward
- Decomposition of long tasks into locally-solvable subproblems, which is close to what domain decomposition does for PINNs
None of this is new advice. What SciML adds is evidence about the cost of the alternative, measured in a domain where you can check the answer against physical law instead of a vibes-based rubric. That last part matters more than it sounds. Electromagnetics has ground truth. Agent evaluation largely doesn’t, which is exactly why we should pay attention to fields that can tell us which methodological instincts actually hold up.
The part I’d hold loosely
One Spotlight paper does not settle a research program, and I don’t yet know how FLASH-MAX generalizes beyond the electromagnetic setting or what it costs to train. Maxwell’s equations are an unusually good case for structural priors: exactly known, linear, and centuries-refined. Very little in agent work comes with a prior that clean.
Still, the direction of travel is clear enough. The systems that hold up are the ones where correctness is a property of the construction, not a request made politely at inference time. Physics got there first because physics keeps score.
🕒 Published: