We hear of breakthroughs nearly every day driven by a wide variety of various applications that leverage deep learning techniques; from image and voice recognition to fraud detection and diagnostic medicine. Perhaps the most well-known examples where machine learning is pervasive is teaching drones to navigate and perform tasks, and of course modern autonomous cars that can drive themselves safely in traffic. All of these types of applications have one thing in common; the systems are relatively easy to understand, describe, model and deploy and there is also a wealth of data with which to train. However in contrast, the world of organic chemistry deals with molecular structures – and the connections between a given structure, its many properties, and all of the potential applications and uses is a very complex relationship. Designing new molecules that exhibit a desired set of properties remains the holy grail of chemistry – and one researcher at Harvard University is trying to teach computers how to identify them.
Dr. Alan Aspuru-Guzik, Professor of Chemistry and Chemical Biology has been exploring the world of high performance and distributed computing for quite some time. The Clean Energy Project originated in his group and took the form of a screen saver that was designed to help identify materials to design the next generation of low cost, high efficiency solar cells and increase the output of this critical renewable energy resource. Imagine, solar cells as inexpensive to manufacture as a plastic bag! Screening millions of candidate molecules to identify candidates with favorable excitation energy transfer properties is a daunting task, but Alan’s group tapped into the World Community grid with special code that used the available GPU cycles on gaming systems around the world. They developed a system that uses a high throughput virtual discovery and design process to reduce the pool of candidates to those which have potential. To understand why so much processing is necessary, we can imagine the scope of an unstructured space such as molecules with drug-like properties for example, which has been estimated to be more than 1052 times larger than the total number of compounds ever synthesized! Once a set of candidates have been isolated, however, rigorous analysis of the energy properties at the needed level of accuracy of a single molecule can take literally billions of CPU-hours.
Today, with the potential of artificial intelligence techniques for complex analysis and discovery, the group is turning a new focus on developing an ecosystem of descriptive spaces, datasets, frameworks, and properties for chemistry that can work within a deep learning framework. Unlike images or languages, the three dimensional world of physical, orbital, electrostatic, and quantum space in which chemical molecules exist and interact, has no common, universal schema for representation. Developing a data-driven continuous representation of discrete molecules that can be understood and manipulated by neural networks is one of the first challenges that the group tackled.
Looking to the future, any number of applications using chemical compound design and property prediction could potentially benefit from deep learning techniques. Alan’s group remains active in the area of research exploring different models and training methods to make the best use of this technology, but when they need to scale, they benefit from the University’s HPC cluster, Odyssey, and the Research Computing group who manages the system. Since it was first launched back in 2007, Odyssey has grown into an InfiniBand-connected cluster in excess of 60,000 cores, with more than 35PB of storage comprised of multiple classes of filesystems, each support different performance and workflows. The cluster supports several schools within the Harvard University system, touching nearly every area of technology and research. A diverse team of scientists, engineers, and administrators who comprise the group pride themselves as, “enablers of scale for scientists,” not only managing a cluster but helping to translate scientific challenges to high performance cluster computing solutions. To quote Scott Yockel, Interim Director of Research Computing at Harvard, “We love when we help researchers have an epiphany that they are simply holding a problem the wrong way. People tend to think serially, but HPC clusters excel at parallel tasks.”
Time, as the saying goes, is money. This is as true in the commercial sector as it is in academic research. The need to accelerate and drive research in a timely and cost-effective manner has never been more needed than it is now. As the Odyssey system grows and demand shifts to more compute-intensive machine learning and hybrid tasks, the in-network computing capabilities of Mellanox InfiniBand such as Scalable Hierarchical Aggregation and Reduction Protocol (SHARP)™, GPUDirect™ and of course RDMA, hold the promise of accelerating the machine learning and HPC workloads central to this research.