The robotic kitchen staff was engrossed in their culinary tasks, diligently working in a compact space filled with advanced equipment. One mechanical arm deftly selected and blended ingredients, while another smoothly maneuvered along a fixed track, managing the ovens. A third arm was responsible for plating, skillfully transferring contents from a container onto a dish. Gerbrand Ceder, a renowned materials scientist from Lawrence Berkeley Lab and UC Berkeley, observed with satisfaction as a robotic arm delicately sealed an empty plastic vial—a particularly intricate task that he enjoyed witnessing. “These machines can operate tirelessly,” remarked Ceder, sharing a knowing glance with two of his graduate students.
Housing materials such as nickel oxide and lithium carbonate, the A-Lab facility was specifically crafted to innovate and produce novel materials, with a focus on potential applications in future battery technologies. The outcomes of their experiments often proved to be unpredictable. Just like a human scientist, the robots sometimes succeeded in creating a fine powder, while other times the result was a sticky, melted mess or a complete evaporation of the material. “In such cases, humans would need to decide the next course of action,” explained Ceder.
The robotic systems were designed to emulate this process. They would analyze their creations, tweak the recipes, and repeat the cycle. “You provide them with instructions in the morning, and by the time you return, you might be greeted with a freshly baked soufflé,” mentioned materials scientist Kristin Persson, Ceder’s close research partner at LBL and also his spouse. However, the outcome could also be a burnt disaster. “Nevertheless, the next day, they will undoubtedly produce a much-improved dish.”
The repertoire of recipes available to Ceder’s robots expanded significantly recently, courtesy of an AI program developed by Google DeepMind known as GNoME. By leveraging data from the Materials Project, a comprehensive database of 150,000 known materials overseen by Persson, the AI system generated designs for 2.2 million new crystals, out of which 380,000 were deemed stable—unlikely to decompose or pose hazards, making them prime candidates for laboratory synthesis. This substantial increase in stable materials opens up possibilities for advancements in solid-state electrolytes, solar cell components, or high-temperature superconductors, as highlighted in a recent publication in Nature.
The quest to identify these potential breakthroughs begins with the synthesis process, underscoring the importance of working efficiently, even through the night. In a recent series of experiments at LBL, as detailed in another Nature publication, Ceder’s autonomous lab successfully fabricated 41 of the materials theorized by GNoME over a span of 17 days. This achievement not only validated the AI model but also showcased the effectiveness of the lab’s robotic methodologies.
Determining the feasibility of creating a material, whether through manual labor or robotic assistance, often hinges on its stability. Essentially, the atoms in the crystal structure must be arranged in the most energetically favorable state to ensure stability. The evolution of stable materials has evolved over millennia, initially through natural discoveries or serendipitous findings, and more recently, through computational design.
However, Persson pointed out a prevalent challenge in this domain—the tendency towards bias. The accumulated knowledge in the field tends to favor familiar structures and elements, akin to the “Edison effect,” where Thomas Edison’s relentless experimentation ultimately led to the discovery of the optimal lightbulb filament material after numerous trials. To counteract these biases, DeepMind’s GNoME takes a fresh approach by leveraging active learning techniques to explore novel material designs beyond conventional boundaries.
By employing a graph neural network (GNN) to identify patterns in stable structures and employing quantum mechanics principles like density-functional theory (DFT) to refine the results, GNoME can propose complex structures involving multiple elements, surpassing the limitations of human intuition. The iterative process of generating, verifying, and adjusting material designs has proven instrumental in expanding the realm of possibilities in material science.
While the theoretical validations through DFT are crucial, the ultimate test lies in the physical synthesis of these materials. Ceder’s team selected 58 crystals from the theoretical pool for fabrication in the A-Lab, acknowledging the initial failures that prompted iterative adjustments in the robotic recipes. After the rigorous experimentation period, the lab successfully produced 41 materials, demonstrating the potential of AI-driven material discovery and synthesis methodologies.
While the prospects of automating material synthesis are promising, some experts caution against the impracticality of blindly pursuing thousands of hypothetical materials without specific objectives. The balance between theoretical predictions and practical synthesis remains a key challenge in material science research. Despite the vast array of materials proposed by GNoME, only a fraction may prove viable for real-world applications, considering factors like material availability, synthesis feasibility, and commercial viability.
The journey from conceptualizing a new crystal to translating it into a marketable product involves numerous hurdles, including compatibility testing, performance evaluation, and scalability considerations. The collaborative efforts between AI-driven material design and human expertise aim to push the boundaries of material science and inspire innovation in diverse applications.
As the field embraces the influx of new materials and data, researchers are poised to explore uncharted territories, driven by the quest for groundbreaking discoveries. The convergence of AI capabilities, advanced synthesis techniques, and interdisciplinary collaborations heralds a new era of material science, offering boundless opportunities for scientific exploration and technological advancement.
