For materials groups, it would be better to start with code fragments around existing simulators such as DFT, 분자동역학, 상장/상분리 phase-field, and FEM. For example, the agent could be asked to revise an input parser, a boundary-condition generator, a parameter-sweep logic, or a surrogate subroutine candidate, and then the outputs of the simulator could be ranked by physical metrics.
Next, it is safer to score “feasible answers that violate physical constraints less” rather than “better answers.” Rather than inventing a new model form, a suitable pilot would reconstruct existing empirical rules or heuristics in code and then examine how much the simulation results improve them.