A-Lab: Pioneering Autonomous Materials Discovery Through AI and Experimentation

By Dr Silpaja Chandrasekar, PhDReviewed by Susha Cheriyedath, M.Sc.Dec 3 2023

In a paper published in the journal Nature, researchers detailed the A-Lab, an autonomous laboratory pioneering the fusion of computational screening and experimental material creation. Over 17 days, employing artificial intelligence (AI), robotics, and historical data, the lab crafted 41 new compounds from 58 targets. The study of failed syntheses offered actionable insights for refining material screening and synthesis methodologies. This success underscores AI's pivotal role in discovering autonomous materials, advocating for the further amalgamation of technology and experimental endeavors.

Background

The discovery of new materials through computational methods often outpaces their experimental realization due to the challenges and time involved in the latter. There's a growing need for autonomy in experimental agents to expedite this process, enabling the interpretation of data and decision-making.

Previous strides in materials research showcased autonomy in various aspects, employing robotics and Bayesian-driven optimization for tasks like carbon nanotube yield, photovoltaic performance, and photocatalysis activity. While conventional machine learning (ML) aids optimization, human researchers draw upon extensive background knowledge, highlighting the necessity of integrating domain knowledge, diverse data access, and active learning for autonomous systems.

Fundamental Processes in A-Lab Workflow

Materials Screening: The A-Lab identified 58 target materials from the Materials Project database, focusing on theoretically stable compounds absent in the Inorganic Crystal Structure Database (ICSD) and excluding certain elements (radioactive, rare, toxic). Selection was limited to specific material types due to handling concerns. Candidate materials underwent novelty verification using Synergistic Earth and Terrain Resources Alliance (SynTERRA) and the 'Handbook of Inorganic Substances,' filtering down to 146 new compounds stable in air. From these, 58 targets with readily available precursors were chosen, with most believed to have no prior reports.

Synthesis Recipe Generation: A synthesis pipeline recommended recipes based on a knowledge base of solid-state procedures, employing initial selection and similarity-based strategies. Using precursor properties, target composition, and thermodynamic driving forces, ML models predicted effective synthesis temperatures. The lab averaged fixed temperatures from the recommended precursor sets to enhance operational efficiency.

Robotic Synthesis and Characterization: The A-Lab employed an autonomous automated platform for solid-state synthesis and characterization, including precursor preparation, mixing, heating, and X-ray diffraction analysis. The process involved various stations and robotic arms for sequential handling and testing.

Phase Analysis: The lab used X-ray diffraction (XRD) AutoAnalyzer, relying on Convolutional Neural Network (CNN)-based phase identification and estimation of weight fractions from XRD patterns. An automated Rietveld refinement approach further validated predictions, and manual analysis was conducted when necessary, particularly for cases of configurational disorder.

Active Learning Algorithm:  Advanced Rotorcraft Rotor Blade Optimization and Windfarm Load Simulation for 3D (ARROWS3), an active-learning algorithm, iteratively explored reaction pathways by adjusting temperatures, recording outcomes, and proposing new experiments. It prioritized precursor sets likely to avoid unwanted reactions and sought to improve target yield by identifying optimal reaction paths.

Experimental Synthesis and Computational Insights

Experimental Synthesis Outcomes: Employing its workflow, the A-Lab achieved a 71% success rate, synthesizing 41 out of 58 target compounds over 17 days. Modifying the lab's decision-making algorithm and computational techniques could raise this success rate to 78%. A graphical representation shows the outcomes of the synthesized compounds against their decomposition energies, highlighting stability and metastability trends among the targets. Across the tested range, the synthesis success did not exhibit a clear correlation with decomposition energy.

Synthesis Recipe Insights: ML models trained on literature data successfully obtained 35 synthesized materials. These recipes were more successful when reference materials closely resembled the targets, indicating the importance of precursor selection in determining synthesis outcomes. However, the synthesis pathways produced only 37% of the desired compounds, highlighting the substantial influence of precursor selection on the outcomes despite achieving 71% of the target materials.

Active Learning Contributions: The A-Lab's active learning cycle optimized routes for nine targets, six of which had initially yielded zero success with literature-inspired recipes. The lab identified more efficient synthesis routes by prioritizing precursor sets and avoiding unfavorable reaction pathways. Notably, the A-Lab built a database of pairwise reactions, reducing the search space for synthesis recipes and enabling a more thoughtful selection of precursor sets for future experiments.

Identified Barriers to Synthesis: The researchers could not synthesize 17 of the 58 evaluated targets. These failures stemmed from four main categories: slow reaction kinetics, precursor volatility, product amorphization, and computational inaccuracies. Strategies to overcome these barriers include adjusting synthesis parameters, refining precursor selection algorithms, and addressing challenges in computational predictions.

Challenges and Computational Insights: Issues with computation inaccuracies impacted a few targets, revealing errors in predicted energies for specific compounds. For instance, challenges in predicting the stability of compounds like La₅Mn₅O₁₆ were attributed to fundamental electronic structure difficulties, highlighting the A-Lab's role in providing feedback to enhance computational accuracy for better predictive modeling.

Conclusion

To sum up, the A-Lab's successful synthesis of 41 out of 58 target compounds demonstrates the power of integrated computational modeling and active learning in materials discovery. Despite encountering barriers such as slow reaction kinetics and computational inaccuracies, the insights gained underscore the importance of refined synthesis strategies and improved computational methodologies in advancing materials research and discovery.