Research Notes

Mar 24, 2026

Lab update: automated energy discovery stack

A look at the orchestration layer we use to generate hypotheses, run simulations, score outputs, and decide what should be explored next.

Closed-Loop Discovery

The discovery stack links hypothesis generation, simulation execution, scoring, and follow-up selection into one loop. The loop is not meant to be autonomous in the theatrical sense. It is meant to make research state explicit: what was tried, why it was tried, what failed, and which constraint changed the next decision.

Recent materials work has shown that AI-generated candidates and autonomous synthesis can be connected into a productive workflow. That does not remove the need for scientific judgment. It changes where the judgment sits. Researchers spend less time hand-enumerating obvious candidates and more time defining constraints, failure modes, and evidence thresholds.

Why Energy Research Needs Orchestration

Energy materials and systems are multi-objective by default. A solid electrolyte candidate is not useful only because it is stable in a database. It also has to be synthesizable, conductive enough, chemically compatible, mechanically workable, and safe. A grid-control policy is not useful only because it optimizes a simulated objective. It has to be robust to bad forecasts, missing telemetry, and operator constraints.

The stack therefore treats every candidate as a bundle of linked claims. One model may estimate stability, another may estimate manufacturability, another may score operational value, and a human reviewer may flag a safety or supply-chain issue. The output is not a single ranking. It is a structured case for why the candidate deserves the next experiment.

From Candidate Generation To Experimental Queue

The candidate generator proposes materials, operating policies, or nuclear-host configurations. The simulator tests them against fast approximations. The critic models look for constraint violations and uncertainty hotspots. The planner selects the next batch so the lab learns as much as possible from a limited number of runs.

This is where automated discovery becomes practical. The system should sometimes choose an unglamorous baseline because it reduces uncertainty. It should sometimes choose an apparently weak candidate because it tests whether a model is wrong. A lab that only tests the top predicted score is not really learning.

What We Log

Every run records prompts, constraints, model versions, simulation settings, intermediate scores, rejected candidates, and the reason each candidate moved forward or stopped. That history matters because research failures often contain the next useful hypothesis.

The long-term goal is an auditable discovery graph. A researcher should be able to open a result months later and see the chain of evidence, not just the final plot. That is especially important for energy systems, where deployment decisions can be expensive and failure modes are not confined to a screen.

Research basis

2026