Published August 12th, 2021
Santiago Miret is a deep learning researcher at Intel Labs, where he focuses on developing artificial intelligence (AI) solutions and exploring the intersection of AI and the physical sciences.
- Intel Labs creates new research effort to investigate and develop advanced AI algorithms and technologies to accelerate the discovery of new material systems more economically.
- Intel Labs and the Matter Lab at the University of Toronto collaborate to develop novel methods for AI-driven design of molecular materials.
The Economic Opportunities of Materials Design
The successful design and deployment of novel material technologies in the last couple of decades has enabled tremendous innovations across various industries. Building today's smartphones, for example, would have cost about 100 million dollars in the 1980s and yielded a 14 meters tall device, both of which would be very impractical. Since the 1980s, however, there have been tremendous advances in materials technologies that have made the smartphone possible, including:
- Miniaturization of electronics with computer processors packing a lot more computing power on a smaller area thanks to advances in materials design and processing technology primarily driven by Intel and other semiconductor companies.
- Denser and more reliable battery packs based on lithium-ion technology instead of lead-acid batteries.
- Touch screens leveraging novel materials with specific optical and electronic properties, such as indium tin oxide, which make up transparent layers that conduct electricity in order to recognize a finger's touch based on the change in electric charge unique to human skin.
Furthermore, materials innovations surrounding silicon have enabled advances in microelectronics and computer technologies that build the foundation of a technology-enabled world, including the recent proliferation of artificial intelligence (AI). Similar, albeit different advances, in silicon technology and perovskites, a class of semiconductor materials that transport the electric charge of light, have provided the basis for solar photovoltaic cells which enable the harvesting of renewable solar energy thereby driving a redesign of the energy industry to a more sustainable and less carbon-heavy system.
In the pharmaceutical industry, an enormous variety of molecular compounds have been discovered and made into effective drug therapies that continue to push the boundaries of medicine and the diseases we can successfully mitigate and cure. While the diversity and impact of various material technology applications show the immense promise of next-generation materials designs, the cost of discovery, design, and deployment of various material systems has been increasing across some critical industries. In semiconductors, for example, materials advances that underlie process technologies have been getting more expensive.
Figure 1: Cost of new semiconductor technology nodes over time-based on data from McKinsey on Semiconductors, No 1, Autumn 2011.
While Moore's Law continues to drive compute advances, those same advances are getting more and more expensive to realize. The closest parallel to Moore's Law in the pharmaceutical industry is Eroom's Law, which has a more pessimistic observation, unlike Moore's Law. Eroom's law states that drug discovery is becoming slower and more expensive over time, with the cost of bringing a drug to market approximately doubling every two years.
The cost increases observed in Eroom's Law have persisted since the 1980s, despite various improvements in technology. Motivated by these increasing cost trends, along with tremendous advances in AI, our team at Intel Labs decided to start a new research effort to investigate whether advanced AI algorithms and technologies could improve the process for and accelerate the discovery of new material systems more economically.
Partnering with Research Leaders
While the promise of AI-enhanced materials discovery is immense, the underlying technical challenge is equally, if not more, immensely complex. Advanced materials can have many different interacting pieces that can behave in unpredictable ways, creating a large modeling and design challenge. Given the complexity of this challenge, Intel Labs partnered with the Matter Lab led by Alán Aspuru-Guzik, professor of Chemistry and Computer Science at the University of Toronto. Alán is also a past and present research leader at the intersection of chemistry and algorithmic materials design.
Matter Labs’ mission is to accelerate the discovery of new chemicals and materials for novel technologies, including quantum computers, leveraging machine learning and end-to-end automation. The Lab’s research vision centers on automating the materials design, synthesis, characterization and analysis pipeline and create a closed loop of design automation. Our collaboration aims to jointly leverage the expertise of both labs to develop novel methods for AI-driven design of molecular materials and touches the intersection of chemistry, computer science, mathematics, and engineering.
The technical challenges at the core of this collaboration are highly intricate and interconnected. Some of the relevant research questions that our groups are working on include:
- What are good data structures to effectively represent molecular materials?
The current literature offers a variety of approaches. Some choose to represent molecules as string characters (for example, smiles or selfies) that adhere to specific chemically informed grammar. Others prefer a graph-based representation where atoms represent nodes and chemical bonds represent edges in a connected graph. Each representation has distinct advantages and disadvantages, so this question will continue to be on an ongoing research problem.
- What are effective AI methods that can provide novel material designs?
The question of which AI method to use is closely tied to the chosen data representation. In string representations, past research work has leveraged generative adversarial networks (GANs) to provide new string suggestions based on a dataset of strings. This approach, however, not only requires a large dataset, but can often also be limited to generating suggestions that fall within a similar distribution to the training dataset.
- What are some of the alternative ways to evaluate the "goodness" of a given material?
This question is fundamentally tied to the desired application of the underlying material. A material that is a promising drug candidate is unlikely to also perform well for semiconductor applications. Nevertheless, modeling how well a given material will perform in each application setting is challenging and an active area of research in various scientific communities.
- What are effective computational tools that we can leverage to build a successful framework?
Assuming we can design a promising research framework, its success will still depend heavily on good engineering practices that leverage state-of-the-art tools. While there already exists a richness of widely adopted computational chemistry packages by the science community, they have been sparsely used by the AI and deep learning (DL) community. They lack essential features critical for DL training and inference, such as automatic differentiation and the ability to handle tensor-based data. Moreover, research at the intersection of AI and the sciences has received increased attention in recent months and years, leading to a set of highly fragmented toolkits and frameworks that often address a sub-section of a greater problem.
The questions above indicate some of the different components required to create a successful AI framework that can enable effective materials design. The interconnectedness of the various components requires an interdisciplinary team with complementary expertise to ensure that each part is working effectively and connecting them to an operational framework.
Previous Work at Intel Labs
Prior research at Intel Labs has shown promise in addressing real-world problems with interconnected aspects. In a paper published at the International Conference on Learning Representations (ICLR 2021), titled "Optimizing Memory Placement using Evolutionary Graph Reinforcement Learning," we successfully applied a novel reinforcement learning algorithm to provide 28-78 percent speedups of AI workload inference through enhanced memory mappings.
Our recent work, "Neuroevolution-Enhanced Multi-Objective Optimization for Mixed-Precision Quantization," showed success in optimizing AI workload compression while considering multiple objectives. These algorithms, developed by Intel Labs, have demonstrated advancements on systems-level problems, which gives us confidence that we can apply similar ideas to materials design challenges.
We believe that our current research only represents the tip of the iceberg in potential improvements for materials design. Furthermore, once a material design is determined, the problem then shifts to synthesis, fabrication, and characterization, all of which can also be improved by AI-enhanced technologies. Examples of AI technologies used for synthesis include using AI models to improve yield and make chemical reaction predictions; in the case of characterization one can imagine the application of computer vision methods to rapidly analyze microscopy images for various kinds of defects.
We look forward to continuing to work with the Matter Lab on these fascinating, challenging, and highly impactful intersectional research challenges.
You must be a registered user to add a comment. If you've already registered, sign in. Otherwise, register and sign in.