Thermal pathway reconstruction exposes a reliability gap for AI-guided materials design
- 1Stern College for Women
- 2Yeshiva University
Abstract
AI-guided materials design increasingly connects candidate generation with processing and experimental planning, yet recognizing a material’s final thermal endpoint is not the same as reconstructing how it gets there. We benchmarked closed-book reconstruction of source-qualified differential scanning calorimetry event maps for 30 molecular materials using GPT-5.6 Luna, GPT-6 Luna, GPT-5.6 Sol, and Claude Opus 5.5. In the primary API experiment, models recovered terminal melting in 194/216 opportunities (89.8%) but other thermal transitions and pathways in only 105/318 (33.0%). Additional frontier-model configurations reproduced this endpoint-pathway gap. Providing source-reported preparation and scan conditions did not improve pathway reconstruction: recall changed by −6.5 percentage points (95% CI, −14.5 to +0.9) for Sol and −4.9 points (−13.6 to +4.5) for Opus. Frontier models can therefore recognize where a material’s thermal trajectory ends while failing to reconstruct the state changes determining how it gets there.
Related
- Study on the Thermal Properties and Reliability of Liquid Metal/Copper-Coated Diamond Composite Thermal Interface Materials
- Systematic evaluation of LLM-based zero-shot model construction in chemistry and materials science: predictive performance and response reliability
- An overview of aerogels as advanced thermal insulation materials
- Nano-Based Advanced Thermoelectric Design: Materials Engineering and Thermal Management Applications
- Thermal Formation of a Magenta Merocyanine in Ethanol: A Microsolvated SP→TCC→TTC Pathway in a Nitrospiropyran
BibTeX
@misc{lavinda2026thermal,
title = {{Thermal pathway reconstruction exposes a reliability gap for AI-guided materials design}},
author = {Olga Lavinda},
howpublished = {ChemRxiv},
year = {2026},
month = oct,
doi = {10.26434/chemrxiv.15009710/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15009710/v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - Thermal pathway reconstruction exposes a reliability gap for AI-guided materials design
AU - Olga Lavinda
PY - 2026
DA - 2026/10/01
PB - ChemRxiv
DO - 10.26434/chemrxiv.15009710/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15009710/v1
AB - AI-guided materials design increasingly connects candidate generation with processing and experimental planning, yet recognizing a material’s final thermal endpoint is not the same as reconstructing how it gets there. We benchmarked closed-book reconstruction of source-qualified differential scanning calorimetry event maps for 30 molecular materials using GPT-5.6 Luna, GPT-6 Luna, GPT-5.6 Sol, and Claude Opus 5.5. In the primary API experiment, models recovered terminal melting in 194/216 opportunities (89.8%) but other thermal transitions and pathways in only 105/318 (33.0%). Additional frontier-model configurations reproduced this endpoint-pathway gap. Providing source-reported preparation and scan conditions did not improve pathway reconstruction: recall changed by −6.5 percentage points (95% CI, −14.5 to +0.9) for Sol and −4.9 points (−13.6 to +4.5) for Opus. Frontier models can therefore recognize where a material’s thermal trajectory ends while failing to reconstruct the state changes determining how it gets there.
ER -