TRACE stops fixing the molecular neighborhood: a success/failure controller widens or narrows the search, and 80 DFT evaluations lift FePc dissolution resistance to 2.59 eV

In open-ended molecular optimization, the size of the search region is now set by optimization progress rather than fixed before the run. Earlier approaches such as graph genetic algorithms and STONED encoded the neighborhood in the candidate-generation mechanism, with similarity criteria or per-iteration candidate counts specified before optimization and held fixed throughout. Across six PMO objectives under a matched 1,000-evaluation budget (20 initial molecules, then 20 batches of 49), TRACE has a mean top-1 AUC rank of 1.73 and a mean top-10 AUC rank of 1.27, the best among the compared methods.

The search region is split into two adjustable quantities: how far chemically, and how wide the candidate set. A Tanimoto-similarity threshold alone bounds how dissimilar an admitted molecule may be from the current search center; it does not determine how many candidates reach the acquisition function. The threshold τ and the cardinality budget k are coupled by a monotone map, so as k moves between kmin and kmax, τ falls from τhi to τlo, giving one adaptive state that controls both.

The controller only expands or contracts the region from evaluation outcomes; it does not pick individual molecules. Candidates come from a graph-perturbation mechanism that never reads objective values, so objective information previously entered only at the surrogate and acquisition stage. After nsucc consecutive successes k doubles, after nfail consecutive failures it halves, and a contraction below the feasible batch floor triggers a restart that resets k; the GP and qLogEI then select batches only within the admissible set.

Under expensive first-principles evaluation, the search still travels far through local steps. The initial FePc reference has a dissolution energy of 0.42 eV and is not counted among the subsequent 80 designed-candidate evaluations. After eight rounds of 10 DFT evaluations each, the best structure has Ediss = 2.59 eV; it has Tanimoto similarity 0.65 to the round-8 search center, which is itself about 0.19 similar to the initial FePc.

The demonstrated scope is a low-budget benchmark on six PMO objectives, a crossed ablation on two objectives, and a single closed-loop FePc DFT campaign on one objective; the authors frame the latter as an expensive-oracle demonstration rather than a comparison of optimizers. It suits readers interested in small-budget molecular optimization and AI-guided materials design. The authors propose a more systematic understanding of how candidate generation interacts with adaptive region control as the next step.

The authors note that the ablation diagnostics cannot explain why region control helps only with the TRACE proposer, and the conclusion is restricted to the two objectives and five seeds studied. The FePc study is a single computational campaign on one objective, and practical catalyst design will require handling competing objectives and constraints together with experimental validation. Fingerprint similarity is only an imperfect proxy for objective locality, and the linear coupling between τ and k is one of several possible designs.