Anchored Scenario Coverage for Failure-Aware First-Hit Batch Inverse Design

Chuhan Yang; Chenxi Wang; Linhan Wu; Yuyang Liu

math.OCAug 28, 2026

Anchored Scenario Coverage for Failure-Aware First-Hit Batch Inverse Design

Authors: Chuhan Yang, Chenxi Wang, Linhan Wu, Yuyang Liu

Organizations: DeepVerse

Abstract

Early discovery of at least one valid design satisfying a target requirement is a central objective in failure-prone closed-loop inverse design. A natural batch baseline ranks candidates by a product-form marginal valid-hit score, but selecting the highest-ranked candidates independently can produce redundant recommendations under predictive uncertainty and waste the experiment budget. We introduce ARC-SC(Anchored Risk-Constrained Scenario Coverage), a batch acquisition method that preserves strong marginal candidates as anchors and allocates the remaining batch positions by maximizing complementary coverage over predictive target scenarios under a risk-support constraint. In frozen-oracle closed-loop simulations on superconductivity and JARVIS materials-property benchmarks, ARC-SC yields a statistically supported improvement in first-hit discovery and remains competitive with directionally favorable first-hit performance on more challenging design space. These results establish ARC-SC as a POF-anchored, scenario-aware batch strategy for improving early valid-target discovery under structured experimental failure.

Figures & tables

Panel of Figure 1 — Figure 1: Integrated schematic overview of ARC-SC. At each closed-loop iteration, target and validity models are updated from previous valid and failed experiments and used to score a finite candidate pool. Strong marginal valid-hit candidates are retained through POF-style risk-admissible anchoring. Multiple predictive target scenarios then expose alternative plausible opportunities, and the remaining batch positions are assigned to complementary candidates under a validity-support risk budget. Experimental evaluation reveals valid target hits, valid non-hits, and invalid outcomes, which are appended to the observed data for the next iteration. The embedded design-space illustrations are schematic and do not represent a particular experimental run.

DATASET	SAMPLES	FEATURES	THRESHOLD	ORACLE $R^{2}$	SPEARMAN $\rho$
Superconductivity	21,263	5	80.0	0.913	0.956
JARVIS CFID	55,723	15	1.539	0.838	0.912

Table 1: Summary of the frozen-oracle benchmark environments. The JARVIS response is negative formation energy per atom, such that larger values are preferred.

Dataset	W/T/L	Cond. win frac.	Sign $p$	$\Delta\widetilde{\tau}_{\mathrm{hit}}$
Superconductivity	70/251/39	$0.642\,[0.545,0.732]$	$0.004$	$-0.269\,[-0.539,-0.025]$
JARVIS CFID	49/272/39	$0.557\,[0.447,0.663]$	$0.337$	$-0.117\,[-0.297,0.053]$

Table 2: First-hit comparison against the POF baseline using ten seeds per condition. W/T/L denotes earlier-hit wins, ties, and later-hit losses. The conditional win fraction is computed over discordant pairs only.

\Delta\widetilde{\tau}_{\mathrm{hit}}<0

favor ARC-SC. Intervals are

95\%

confidence intervals.

Figure 2: First-hit probability as a function of experimental round. Lines show ARC-SC and POF baseline; shaded regions denote hierarchical-bootstrap

95\%

confidence intervals. Left: superconductivity with shared uniform candidate pools. Right: JARVIS with the mixed candidate-pool protocol. The results show that ARC-SC improves early target discovery on superconductivity with statistically significant gains in first-hit performance. On JARVIS, ARC-SC remains competitive. Although it does not achieve statistically significant superiority, it demonstrates robustness in the more challenging high-dimensional design space.

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Quantity	Superconductivity	JARVIS CFID
Initial observations $n_{0}$	50	50
Batch size $B$ / rounds	5 / 5	5 / 5
POF anchors $K$	3	3
Candidate-pool size $N$	10,000	10,000
Predictive scenarios $M$	64	64
Target quantile	0.85	0.75

Appendix

Table 3: Algorithm and evaluation settings. Dataset-specific entries are shown explicitly.

Model	$R^{2}$	RMSE	MAE	Spearman $\rho$
Extra Trees	0.9128	10.0184	5.6564	0.9556
Random forest	0.9095	10.2057	5.8667	0.9535
Hist. gradient boosting	0.8932	11.0899	6.8611	0.9352

Appendix

Table 4: Held-out superconductivity oracle performance.

Model	$R^{2}$	RMSE	MAE	Spearman $\rho$
Random forest	0.8375	0.4320	0.2518	0.9121
Extra Trees	0.8311	0.4404	0.2531	0.9088
Hist. gradient boosting	0.8239	0.4498	0.2828	0.9002

Appendix

Table 5: Held-out JARVIS oracle performance.

Explore similar work

SciDesignBench: Benchmarking and Improving Language Models for Scientific Inverse Design
Mar 13, 2026David van Dijk, Ivan VrkicInverse Design Large Language Model-Guided Optimization
CellType Inc.
Range-Aware Bayesian Optimization for Discovering Diverse Designs within Target Property Windows
Jun 10, 2026Shengli Jiang, Jason Wu, Charles M. Schroeder +1Molecular Optimization Bayesian Optimization
Department of Chemical and Biological Engineering, Princeton University, Princeton, NJ 08540, USA
Many Needles in a Haystack: Active Hit Discovery for Perturbation Experiments
May 11, 2026Andrea Rubbi, Arpit Merchant, Samuel Ogden +4Bayesian Optimization Bayesian Experimental Design
Department of Computer Science and Technology, University of Cambridge, Cambridge, UK · Cambridge Center for AI in Medicine, University of Cambridge, Cambridge, UK · MediaTek Research, Cambridge, UK +1

Mar 13, 2026J/K move · Enter open · S save

SciDesignBench: Benchmarking and Improving Language Models for Scientific Inverse Design

David van Dijk, Ivan Vrkic

CellType Inc.

Inverse Design Large Language Model-Guided Optimization Scientific Reasoning in Language Models Benchmark Design Fine-Tuning

Open paper AskPDF arXiv

Many of the most important problems in science and engineering are inverse problems: given a desired outcome, find a design that achieves it. Evaluating whether a candidate meets the spec is often routine; a binding energy can be computed, a reactor yield simulated, a pharmacokinetic profile predicted. But searching a combinatorial design space for inputs that satisfy those targets is fundamentally harder. We introduce SciDesignBench, a benchmark of 520 simulator-grounded tasks across 14 scientific domains and five settings spanning single-shot design, short-horizon feedback, long-horizon refinement, and seed-design optimization. On the 10-domain shared-core subset, the best zero-shot model reaches only 29.0% success despite substantially higher parse rates. Simulator feedback helps, but the leaderboard changes with horizon: Sonnet 4.5 is strongest in one-turn de novo design, whereas Opus 4.6 is strongest after 20 turns of simulator-grounded refinement. Providing a starting seed design reshuffles the leaderboard again, demonstrating that constrained modification requires a fundamentally different capability from unconstrained de novo generation. We then introduce RLSF, a simulator-feedback training recipe. An RLSF-tuned 8B model raises single-turn success rates by 8-17 percentage points across three domains. Together, these results position simulator-grounded inverse design as both a benchmark for scientific reasoning and a practical substrate for amortizing expensive test-time compute into model weights.

Abstract

Explore similar work

SciDesignBench: Benchmarking and Improving Language Models for Scientific Inverse Design

Range-Aware Bayesian Optimization for Discovering Diverse Designs within Target Property Windows

Many Needles in a Haystack: Active Hit Discovery for Perturbation Experiments