[Paper Review] Precursor recommendation for inorganic synthesis by machine learning materials similarity from scientific literature

Tanjin He, Haoyan Huo+4 more|arXiv (Cornell University)|Feb 5, 2023

Machine Learning in Materials Science4 citations

TL;DR

This paper proposes a machine learning framework that recommends precursor materials for inorganic solid-state synthesis by learning chemical similarity from 29,900 literature-derived synthesis recipes. Using a self-supervised neural network to encode material compositions, the method achieves an 82% success rate in recommending viable precursor sets for 2,654 unseen target materials, effectively capturing decades of heuristic synthesis knowledge in a scalable, data-driven form.

ABSTRACT

Synthesis prediction is a key accelerator for the rapid design of advanced materials. However, determining synthesis variables such as the choice of precursor materials is challenging for inorganic materials because the sequence of reactions during heating is not well understood. In this work, we use a knowledge base of 29,900 solid-state synthesis recipes, text-mined from the scientific literature, to automatically learn which precursors to recommend for the synthesis of a novel target material. The data-driven approach learns chemical similarity of materials and refers the synthesis of a new target to precedent synthesis procedures of similar materials, mimicking human synthesis design. When proposing five precursor sets for each of 2,654 unseen test target materials, the recommendation strategy achieves a success rate of at least 82%. Our approach captures decades of heuristic synthesis data in a mathematical form, making it accessible for use in recommendation engines and autonomous laboratories.

Motivation & Objective

• To automate the literature-based synthesis design process used by experimental researchers, who typically search for similar materials to repurpose existing recipes.
• To learn chemical similarity between inorganic materials from a large corpus of text-mined synthesis recipes, enabling data-driven precursor recommendation.
• To develop a scalable, generalizable method that captures decades of heuristic synthesis knowledge in a mathematical, reusable form for use in autonomous labs and recommendation systems.
• To evaluate the method on a large, historically split test set of 2,654 novel target materials, ensuring robustness and generalizability beyond case studies.

Proposed method

• A self-supervised neural network (PrecursorSelector encoding) is trained to learn vectorized representations of materials based on their precursor compositions, enabling quantification of material similarity.
• Material similarity is computed using cosine similarity between learned embeddings, allowing the system to identify precedent materials with similar synthesis pathways.
• The precursor recommendation strategy selects the most similar known material from a knowledge base and recommends its precursors for a novel target, mimicking human analogical reasoning.
• The model is trained as a multi-task neural network with adaptive loss weighting to jointly optimize for precursor selection and material representation learning.
• Baseline models include 'Most frequent' (empirical probability), 'Magpie encoding' (132-elemental property statistics), 'FastText encoding' (100D word embeddings from materials texts), and 'Raw composition' (83D elemental fraction vectors).
• Data is split by publication year (training: ≤2014, validation: 2015–2016, test: 2017–2020), with prototype formulas used to prevent data leakage from repeated material reports.

Experimental results

Research questions

• RQ1 Can machine learning models effectively learn chemical similarity from literature-mined synthesis recipes to guide precursor selection for novel inorganic materials?
• RQ2 How does a neural network-based representation of precursor compositions compare to traditional chemical encoding methods (e.g., Magpie, FastText, raw composition) in precursor recommendation accuracy?
• RQ3 To what extent can a data-driven, analogy-based approach replicate the heuristic reasoning used by experimental chemists in precursor selection?
• RQ4 What is the success rate of precursor recommendations on a large, historically unseen test set of 2,654 novel materials?

Key findings

• The proposed method achieves a minimum success rate of 82% in recommending viable precursor sets for 2,654 unseen target materials, significantly outperforming baseline models.
• The PrecursorSelector encoding method, which learns from precursor compositions, outperforms traditional encoding methods such as Magpie, FastText, and raw composition in precursor recommendation accuracy.
• The model successfully generalizes to novel materials not seen during training, as validated on a test set of 1,992 materials published between 2017 and 2020.
• The use of self-supervised learning on precursor compositions enables the model to capture complex, non-linear relationships in synthesis knowledge that are not apparent from elemental properties alone.
• The multi-task learning framework with adaptive loss weighting effectively balances precursor selection and representation learning, improving overall recommendation performance.
• The approach effectively encodes decades of heuristic synthesis knowledge from scientific literature into a mathematical form accessible for AI-driven synthesis planning and autonomous laboratory systems.