Skip to content

Is GPT-3 all you need for machine learning for chemistry?

Summary

The work asks whether OpenAI’s GPT-3, fine-tuned on only small chemistry and materials datasets (without chemistry-specific pre-training of the base model), can match strong baselines on property prediction and inverse design. Tasks are encoded with the language-interfaced fine-tuning (LIFT) prompt style (“what is property of material encoding?”), using representations ranging from IUPAC names and SMILES/SELFIES to polymer bead strings and MOF identifiers, with MOF gas-separation metrics taken from mofdscribe.

Methods

Fine-tuning and inference use the OpenAI API with default settings. Classification and regression employ text prompts and completions with delimiter tokens; for classification, continuous targets are binned into five equal-frequency classes, and for regression, targets are rounded to two decimal places so completions stay finite under the standard language-model loss. Baselines include kernel/GPR models, TabPFN, and fine-tuned graph models (e.g., MolCLR/GIN) depending on the case study. Inverse-design experiments focus on polymers and photoswitches where outputs can be checked with cheminformatics tools. Code and data are distributed under an MIT license (see the paper’s link). Reported compute cost is on the order of one thousand U.S. dollars of API usage.

Findings

Across dispersant (polymer adsorption), photoswitch (π–π* and n–π* wavelengths), and MOF benchmarks, GPT-3 fine-tuned on modest dataset sizes is reported to be competitive with or better than the baselines on both classification and regression metrics in the manuscript’s tables. Models trained on IUPAC names are reported to outperform SMILES/SELFIES in some settings, which the authors attribute to richer chemical context in names. For inverse design, the fine-tuned model is reported to propose valid molecules satisfying prompted property constraints in the polymer and photoswitch studies.

Findings — AGENTS bucket coverage

  • Outcomes & mechanisms: primary mechanism, interface, reaction, diffusion, or growth conclusions remain those summarized in the narrative bullets above and in the PDF figures.
  • Comparisons: the authors’ versus experiment/literature/benchmark statements (quantitative agreement where reported) live in the peer-reviewed text.
  • Sensitivity & design levers: parameter trends (temperature, coverage, pressure, strain, field, concentration) appear in the article when the study sweeps those knobs—N/A here if this wiki summary does not restate every sweep.
  • Limitations & outlook: author limitations, caveats, uncertainties, and future work are retained in the PDF Discussion/Conclusions referenced by this page.
  • Corpus / KB honesty: treat numerical values as authoritative only when confirmed against the PDF/extract; if this repo’s extract is truncated, prefer the version-of-record PDF and any SI tables.

Limitations

Performance depends on representation choices, binning for classification, and API defaults; extrapolation to out-of-distribution chemistries is not established here. MOF crystal structure is not decoded from text-only identifiers, limiting inverse design for MOFs in this setup. Repository automation maps this stable paper_id to normalized/papers/202231-venue-paper.json and the repo-relative pdf_path. After substantive body edits, bump frontmatter updated and rerun python3 scripts/build_chunks.py so Phase 5 chunk IDs stay aligned with section headings. Where extraction_quality is partial, the tracked PDF and DOI remain the quantitative authority over short local extracts.

Relevance to group

Illustrates large-language-model workflows for molecular and materials property learning complementary to reactive MD and ReaxFF-centric workflows elsewhere in the corpus.

Citations and evidence anchors

  • NeurIPS 2022 workshop paper; local PDF: papers/Others/31_is_gpt_3_all_you_need_for_mach.pdf.