[Paper Review] Toward Human-AI Co-creation to Accelerate Material Discovery

Dmitry Yu. Zubarev, Carlos Raoni Mendes+6 more|arXiv (Cornell University)|Nov 5, 2022

Machine Learning in Materials Science4 citations

TL;DR

This paper proposes a human-AI co-creation workbench framework to accelerate materials discovery by integrating expert knowledge with AI-driven tools across four core activities: generative modeling, dataset triage, molecule adjudication, and risk assessment. The system enables iterative, knowledge-rich collaboration between subject matter experts (SMEs) and AI agents through a structured, web-based interface that captures and reuses expert judgment, improving discovery efficiency and reducing opportunity costs.

ABSTRACT

There is an increasing need in our society to achieve faster advances in Science to tackle urgent problems, such as climate changes, environmental hazards, sustainable energy systems, pandemics, among others. In certain domains like chemistry, scientific discovery carries the extra burden of assessing risks of the proposed novel solutions before moving to the experimental stage. Despite several recent advances in Machine Learning and AI to address some of these challenges, there is still a gap in technologies to support end-to-end discovery applications, integrating the myriad of available technologies into a coherent, orchestrated, yet flexible discovery process. Such applications need to handle complex knowledge management at scale, enabling knowledge consumption and production in a timely and efficient way for subject matter experts (SMEs). Furthermore, the discovery of novel functional materials strongly relies on the development of exploration strategies in the chemical space. For instance, generative models have gained attention within the scientific community due to their ability to generate enormous volumes of novel molecules across material domains. These models exhibit extreme creativity that often translates in low viability of the generated candidates. In this work, we propose a workbench framework that aims at enabling the human-AI co-creation to reduce the time until the first discovery and the opportunity costs involved. This framework relies on a knowledge base with domain and process knowledge, and user-interaction components to acquire knowledge and advise the SMEs. Currently,the framework supports four main activities: generative modeling, dataset triage, molecule adjudication, and risk assessment.

Motivation & Objective

• To address the inefficiencies and bottlenecks in end-to-end materials discovery, particularly the lack of integrated platforms that orchestrate AI and expert knowledge.
• To reduce time-to-discovery and opportunity costs by enabling continuous, collaborative workflows between SMEs and AI agents.
• To systematize knowledge capture and reuse across the discovery lifecycle—from hypothesis generation to conclusion drawing—through a unified software framework.
• To improve the viability of AI-generated materials by embedding expert-driven risk assessments and confidence-likelihood reasoning into the generation process.
• To develop a flexible, extensible platform that supports diverse materials discovery use cases, including polymers, metal-organic frameworks, and biomass-based materials.

Proposed method

• The framework is implemented as a web-based Discovery Workbench (DWb) that integrates multiple AI and expert tools into a single, orchestrated workflow.
• It supports four key activities: generative modeling for novel molecule design, dataset triage for quality filtering, molecule adjudication for expert validation, and risk assessment using confidence-likelihood plots.
• Confidence-likelihood plots constrain POS (probability of success) assessments based on LOK (level of knowledge) to prevent overconfidence in low-knowledge scenarios and enforce high confidence in high-knowledge cases.
• Expert assessments are visualized with similarity-weighted references (blue squares), while peer-review assessments use consensus circles to reduce bias and increase transparency.
• The system tracks and reuses knowledge across the discovery lifecycle, enabling feedback loops where risk assessments inform and calibrate generative models.
• The platform is designed to support incremental knowledge base curation and to enable multi-agent AI coordination with minimal SME intervention.

Experimental results

Research questions

• RQ1 How can a human-AI co-creation framework reduce time-to-discovery and opportunity costs in materials science?
• RQ2 What mechanisms enable effective integration of expert knowledge with AI tools across the discovery lifecycle?
• RQ3 How can confidence-likelihood plots improve the reliability of expert risk assessments in materials discovery?
• RQ4 In what ways can captured expert knowledge be reused to guide and calibrate AI generative models?
• RQ5 What design principles support symbiotic, explainable, and scalable human-AI interaction in scientific discovery?

Key findings

• The Discovery Workbench (DWb) framework enables the construction of end-to-end discovery applications with improved systematization and flexible integration of AI tools.
• Initial deployment of a PAG (polymer, additive, glass) discovery application demonstrated acceleration and enhanced knowledge management through structured workflow orchestration.
• The use of confidence-likelihood plots successfully constrained expert POS (probability of success) assessments based on LOK (level of knowledge), reducing overconfidence in uncertain cases.
• Peer-review assessments using consensus circles helped reduce individual biases and made decision-making more transparent and defensible.
• Expert knowledge, captured via structured assessments and visualized with similarity references, improved consistency and reusability across discovery iterations.
• The framework supports iterative feedback from risk assessments to generative models, increasing the likelihood of producing viable, high-POTENTIAL candidates.