PaperlayerView exact revision on arXiv

Source-linked AI summary

TRACE: Transition-Aware Residual Control for Multi-Objective Materials Discovery

Kang Zhou, Yujia Tong, Yong Tao, Jingling Yuan

arXiv:2608.23631v1cs.AIcond-mat.mtrl-sci

TL;DR

Multi-objective materials agents often lack edit-level evidence about how executable modifications change properties, complicating refinement when objectives compete. TRACE records evaluated parent–edit–child transitions, estimates reusable edit effects, and ranks edits against current residual violations while protecting satisfied objectives. Under approximately matched evaluation counts, TRACE raises Qwen macro-average hit rate from 18.13% to 25.96% over LLEMA.

  • Problem

    Existing LLM-based materials agents mainly retain candidate-level outcomes rather than which executable edits caused observed property changes, limiting evidence for multi-objective refinement.

  • Method

    TRACE records evaluated parent–edit–child transitions, estimates reusable edit effects, and selects local edits according to current constraint residuals while retaining global LLM exploration.

  • Results

    25.96% macro-average hit rate versus 18.13% for LLEMA under approximately matched property-evaluation counts.

  • Takeaways & Limitations

    Across 14 LLEMABench tasks and two LLM backbones, evaluated transitions serve as a useful feedback unit for oracle-limited materials search.

  • Takeaways & Limitations

    TRACE requires an oracle that may be a simulator, learned surrogate, or experimental measurement, and transition guidance is limited to the executable edit space.

Abstract

from arXiv · show

Multi-objective materials discovery with LLM agents is often limited not only by how many candidates can be proposed, but by how effectively each costly property evaluation informs the next search step. Existing agents mainly store evaluated candidates and their scores, so they know which materials succeeded but not which executable edits caused useful property changes. This makes local refinement difficult when objectives compete and an edit that improves one property may damage another. We propose TRACE, a transition-aware residual control framework that treats evaluated edits as the basic unit of feedback. TRACE records each local refinement as a parent-edit-child transition with observed property deltas, aggregates transition evidence to estimate reusable edit effects, and ranks future edits by their predicted ability to reduce the current candidate's remaining constraint violations while avoiding damage to already satisfied objectives. In a controlled same-backbone comparison, TRACE improves over LLEMA, the state-of-the-art LLM-agent baseline, raising macro-average hit rate from 18.13\% to 25.96\%.

1 INTRODUCTION

TRACE reframes feedback in multi-objective materials discovery around evaluated edits, linking executable modifications to property changes and current constraint violations. It combines transition-aware local refinement with global LLM exploration and improves controlled performance over LLEMA.

  • Existing LLM materials agents mainly retain evaluated candidates and scores, not which executable modifications produced observed property changes.
  • TRACE records parent–edit–child transitions with observed property changes and reuses them to estimate effects of executable edits.
  • Residual-aware edit selection prioritizes modifications addressing unsatisfied objectives while avoiding degradation of properties that already meet targets.
  • TRACE combines feedback-driven local editing with global LLM exploration to reuse observed edit effects while exploring new material-space regions.
  • 25.96% macro-average hit rate versus 18.13% for LLEMA under approximately matched property-evaluation counts.
  • The evaluation covers 14 multi-objective LLEMABench tasks, controlled budgets, component ablations, exploration-allocation studies, and transition-level analyses.

2 PROBLEM FORMULATION

The problem formulation models materials as candidates evaluated by an oracle against multiple property constraints. Search can either propose candidates globally or apply executable edits to evaluated parents, with transitions recording the resulting property changes and residual progress.

  • 2.1 ORACLE-LIMITED MULTI-OBJECTIVE MATERIALS DISCOVERY: A candidate material x belongs to a material space X, and an oracle O returns its task-relevant property vector.
  • 2.1 ORACLE-LIMITED MULTI-OBJECTIVE MATERIALS DISCOVERY: The oracle may be a simulator, learned surrogate, or experimental measurement, requiring only query access to its outputs.
  • 2.1 ORACLE-LIMITED MULTI-OBJECTIVE MATERIALS DISCOVERY: Each task specifies property constraints with feasible sets representing lower bounds, upper bounds, or intervals.
  • 2.1 ORACLE-LIMITED MULTI-OBJECTIVE MATERIALS DISCOVERY: Normalized violations and total constraint residual place heterogeneous properties on comparable scales for assessing progress toward feasibility.
  • 2.1 ORACLE-LIMITED MULTI-OBJECTIVE MATERIALS DISCOVERY: A candidate is valid when it passes task-specific checks and all property constraints, and the objective is to discover candidates satisfying h(x; C) = 1.
  • 2.2 EXECUTABLE EDITS AND TRANSITION FEEDBACK: The search agent may generate a candidate globally or apply an executable edit e ∈ E(x) to an evaluated parent.
  • 2.2 EXECUTABLE EDITS AND TRANSITION FEEDBACK: Edits must be enumerable, executable, and compatible with the material representation before the constructed child is evaluated by the same oracle.
  • 2.2 EXECUTABLE EDITS AND TRANSITION FEEDBACK: A completed transition records the parent, edit, child, parent and child properties, and property delta ∆y = y(x′) − y(x).

3 TRACE: TRANSITION-AWARE RESIDUAL CONTROL

TRACE combines global LLM exploration with residual-aware local editing and online transition memory. It estimates reusable edit effects from observed transitions, ranks edits by current constraint needs, and allocates attempts between local and global routes.

  • 3.1 TRANSITION MEMORY: TRACE initializes an editable parent pool with LLM proposals, then uses transition memory to guide local Delta edits on evaluated parents.
  • 3.1 TRANSITION MEMORY: Transition memory accumulates evaluated local parent–edit–child transitions online for subsequent edit-effect estimation.
  • 3.1 TRANSITION MEMORY: Each transition is annotated with parent and child constraint states, including satisfied numerical constraints and numerical feasibility.
  • 3.1 TRANSITION MEMORY: Transition utility combines boundary behavior, changes in satisfied constraints, and continuous residual progress, separating property change from current-state value.
  • 3.2 ESTIMATING REUSABLE EDIT EFFECTS: TRACE estimates edit effects with hierarchical backoff from exact signatures to operator classes and then global transition statistics.
  • 3.2 ESTIMATING REUSABLE EDIT EFFECTS: When exact evidence is sparse, pooled operator or global statistics provide estimates; with empty memory, local selection is stochastic.
  • 3.3 RESIDUAL-AWARE EDIT RANKING: For selected parents, TRACE ranks unseen executable edits using historical utility, parent-specific residual progress, and novelty.
  • 3.3 RESIDUAL-AWARE EDIT RANKING: Near the feasibility boundary, TRACE separately rewards progress on violated constraints and penalizes damage to satisfied constraints.

4 EXPERIMENTS

TRACE is evaluated across controlled multi-objective materials-discovery settings, showing improved discovery quality, search-budget efficiency, and transition-guided editing, while performance remains task dependent.

  • Main results: TRACE improves macro-average H.R. from 18.13% to 25.96% with Qwen and from 25.39% to 28.74% with Mistral under approximately matched evaluation counts.Corresponding Stab. averages rise from 5.90% to 8.97% with Qwen and from 11.29% to 12.25% with Mistral.
  • Main results: TRACE improves H.R. over LLEMA on 12 of 14 Qwen tasks and 10 of 14 Mistral tasks, although several task–backbone settings decline.The largest gains occur on Solid-State Electrolytes, while Piezo Energy Harvesters show mixed H.R./Stab. changes.
  • Search-budget efficiency: At 1,000 scheduled attempts, TRACE produces 2.47–4.79× more valid discoveries than LLEMA across four representative tasks.TRACE also lowers formula redundancy on all four tasks and completes more property evaluations within the fixed attempt budget.
  • Component ablation: Full TRACE improves H.R. by 10.04 points and Stab. by 10.23 points over random editing while maintaining a 96.63% G.R.The ablation separates executable editing, transition-history guidance, and residual-aware selection on Solid-State Electrolytes.
  • Global–local allocation: H.R. peaks at 50.19% with a balanced 1:1 LLM:Delta allocation, versus 32.4% for LLM-only and 27.7% for Delta-only exploration.The balanced setting yields 154 structures and 73 unique valid formulas, compared with 83 and 22 for Delta-only exploration.
  • Learned edit effects: Across 882 post-warm-up transitions, Δ-direction accuracy ranges from 61.12% to 76.49% and selection advantage from 54.31% to 64.16%.Both measures exceed their respective random or chance references across the four analyzed tasks.

5 RELATED WORK

Related work spans LLM agents, learning-based inverse design, sequential oracle-limited optimization, and methods that model local transformations more directly.

  • LLM-based agents: LLM agents combine language-based reasoning with scientific tools and external feedback for iterative discovery and have been applied to materials design.Examples include Coscientist, CHEMREASONER, and materials-discovery frameworks such as LLMatDesign.
  • Inverse design: Learning-based inverse-design methods generate materials meeting target specifications using latent representations, diffusion, or conditional generation under chemical and property constraints.Representative systems include CDVAE, DiffCSP, and MatterGen.
  • Oracle-limited optimization: Oracle-limited optimization emphasizes reusing evaluation outcomes, while molecular methods model property-improving transformations or select fragment-level edits.Examples include PMO, Augmented Memory, graph-to-graph translation, Modof, and MARS.

6 CONCLUSION

TRACE concludes that transition-aware residual control can reuse evaluated parent–edit–child changes for local refinement while retaining global LLM exploration in oracle-limited materials search.

  • Conclusion: TRACE records parent–edit–child transitions and uses observed property changes to guide local edits according to current constraint residuals.Global LLM proposals remain available for broader exploration.
  • Conclusion: Across 14 LLEMABench tasks and two LLM backbones, TRACE improves hit rate over LLEMA on most tasks under approximately matched property-evaluation counts.Component and transition-level analyses support transition history as actionable guidance beyond executable editing alone.
  • Search state: TRACE builds an evaluated candidate pool, maintains a fresh-valid frontier for local continuation, and stores transition records separately for edit-effect reuse.The frontier prioritizes newly discovered editable valid candidates, whereas transition memory captures how edits changed properties.
  • Local editing: Local edits use structure-preserving same-group elemental substitutions and pass resulting children through the standard construction and oracle-evaluation pipeline.Only successfully evaluated children contribute observed transitions, and inherited property values are discarded.
  • Guidance activation: TRACE begins with safe unseen substitutions during cold start, then activates learned guidance after a warm-up period and minimum transition count.Predicted effects rank edits but do not replace oracle outcomes; unavailable local capacity is reassigned to global generation.
  • Search procedure: The implementation initializes candidate, transition, and valid-discovery memories before allocating scheduled attempts between global proposals and local edits.Both routes share construction and oracle-evaluation, while evaluated local transitions are appended to transition memory.

B.1 BENCHMARK SCOPE AND MATCHED COMPARISONS

The evaluation uses 14 application-oriented LLEMABench tasks and matched LLEMA–TRACE comparisons with shared backbones, criteria, and approximately equal candidate-evaluation targets.

  • Benchmark scope: The benchmark contains 14 multi-objective materials-discovery tasks with target applications, property constraints, and domain rules for validity.TRACE is evaluated with independently specified search state, transition memory, edit-effect estimation, and residual-aware decision policy.
  • Matched comparisons: Controlled comparisons instantiate LLEMA and TRACE on identical tasks, language-model backbones, and evaluation criteria using Qwen2.5-14B-Instruct and Mistral-Small-3.2-24B-Instruct-2506.Both backbones are 8-bit quantized, and each matched task–backbone setting averages five runs targeting approximately 150 evaluated candidates.

B.2 EVALUATION METRICS AND ATTEMPT ACCOUNTING

The evaluation metrics separate candidate quality conditional on property evaluation from end-to-end performance across all scheduled attempts. This accounting distinguishes constraint satisfaction, stability, generation completion, and scheduled-attempt yield.

  • Evaluation-conditional hit rate counts evaluated candidates satisfying all task constraints, while stability rate additionally requires energy above hull below 0.1 eV/atom.
  • Generation rate uses all scheduled attempts as its denominator, including construction and pre-evaluation failures as misses.
  • Scheduled-attempt yield counts an attempt only when it reaches evaluation and satisfies all task constraints.
  • Construction and pre-evaluation failures affect generation rate and scheduled-attempt yield but are excluded from evaluation-conditional hit and stability denominators.
  • Within-run valid-formula redundancy measures repeated normalized formulas beyond their first occurrence, rather than crystallographic diversity.

C EXPERIMENTAL INSTANTIATION OF TRACE

TRACE combines global LLM exploration with transition-guided local refinement under scheduled-attempt budgets, using shared construction and oracle-evaluation pipelines. Long-horizon analyses attribute its discovery gains to both improved evaluation completion and stronger conditional candidate quality, while repeated transition observations improve edit-direction accuracy.

  • TRACE constructs global proposal context from candidate properties and summaries of executed edits, including support, validity, residual gains, and mean property changes.
  • Table 4 decomposes five-run endpoints after 1,000 scheduled attempts into generation rate, evaluation-conditional hit rate, and scheduled-attempt yield.
  • 1.16–1.27× generation-rate gains and 2.02–3.92× conditional H.R. gains raised scheduled-attempt yield by 2.47–4.79× across tasks.
  • Among 882 post-warm-up transitions, direction accuracy rose from 57.0% with no prior exact-edit observation to 75.0% with at least two observations.
  • From 250 to 1,000 attempts, valid rate increased by 8.14–31.29 points across the four trajectories, while direction accuracy varied by task.

D.3 TASK DEPENDENCE AND APPLICABILITY OF TRANSITION GUIDANCE

Transition guidance is task-dependent: ranking quality alone does not ensure hit-rate gains when valid edited children are scarce. Individual transitions show that an edit’s effect depends on the parent state and remaining constraint residuals, while TRACE also evaluates search diversity through formula repetition.

  • Offline selection advantage exceeded the 50% random reference for every short-run task aggregate, but realized H.R. gains aligned more closely with valid edited-child availability.
  • High-k Dielectrics had 64.16% selection advantage, 9.62% valid rate, and a −1.97-point H.R. difference, whereas two tasks above 60% valid rate improved H.R. by approximately 20 points.
  • The Sr→Mg→CaMgTiO3 edit converted an invalid High-k Dielectrics parent into a valid child by closing residual constraints.
  • Replacing Al with B in MgAl2O4 produced Mg(BO2)2 and increased residual from zero to 2.816, illustrating parent- and residual-conditioned edit effects.
  • TRACE had lower within-run formula repetition than LLEMA on every task, including after shared initial formulas were removed.

E.2 ELEMENTAL COVERAGE AND TASK-ADAPTIVE SEARCH FOCUS

TRACE changes elemental coverage in a task-dependent manner rather than imposing a uniform exploration restriction. It achieves higher conditional hit rate under both reduced and expanded observed-element coverage patterns.

  • At the common 500-iteration checkpoint, elemental occurrences are computed from evaluated candidates and normalized separately for each task–method pair.
  • TRACE observed 33 versus 39 elements for Wide-Bandgap Semiconductors and 33 versus 43 for Toxic-Free Perovskite Oxides relative to LLEMA.
  • TRACE observed 40 versus 36 elements for High-k Dielectrics and 30 versus 29 for Hard, Stiff Ceramics relative to LLEMA.
  • TRACE achieved higher conditional H.R. under both reduced and expanded coverage patterns, supporting task-adaptive search focus rather than a uniform coverage–validity tradeoff.

F DISCUSSION AND LIMITATIONS

TRACE reuses recorded edit–property associations as actionable feedback for subsequent search while retaining global exploration. Its effectiveness may decline when edit effects vary across candidate states, and its current local edit space excludes broader structural modifications.

  • TRACE uses transition history as an intermediate representation between candidate-level feedback and full dynamics modeling.Recorded edit–property associations provide predictive and actionable feedback without learning a full parent-conditioned transition model.
  • Ablations and transition-level analyses show that recorded edit–property associations support subsequent search, while global LLM proposals maintain broader exploration.
  • TRACE’s estimator may be less effective when edit effects vary strongly across candidate states.Parent-specific context and predictive uncertainty are identified as possible improvements for robustness.
  • The current local edit space mainly supports chemistry-constrained, structure-preserving elemental substitutions.Vacancies, lattice changes, and larger structural modifications remain future work.
Loading 2608.23631v1…