With the expanding use of high throughput computations and experiments, chemical and materials science are primed to take advantage of the advancements in machine learning and data sciences to drive the next generation of energy, biomedical, electronics materials and more. This symposium looks to explore state of the art research in experimental and computational chemical and materials science and their interface with data sciences and machine learning through keynote speakers and focused panel discussions on challenges and opportunities in the fields. This event is co-sponsored by the BU College of Engineering, the BU Department of Materials Science & Engineering, and the BU Department of Chemistry.
10:30am – 10:45am ET: Welcome and Opening Remarks
Eric Kolaczyk, Director, Hariri Institute for Computing
10:45am -12:20pm ET: Session 1 – Machine Learning in Experimental Chemical and Materials Science
This session will focus on approaches toward generating robust data for chemical reactions that can be used for ML and autonomous optimization. The speakers will highlight developments in automation of chemical reactions to explore multiple variables and generate high quality data that is applied to algorithms for prediction, optimization, and design.
Connor Coley, Henri Slezynger Career Development Assistant Professor, Chemical Engineering, MIT
- Autonomous development of organic reactions with laboratory robots, data science, and machine learning
- Abstract: Advances in laboratory automation continue to lower the time and expertise required to set up and run chemical reactions; a variety of platforms are now able to screen organic reactions without requiring human intervention. This presents an enormous opportunity not only for high-throughput data generation to train machine learning models, but also for integration with experimental design algorithms. Optimization algorithms can learn from previous experiments to propose new experiments that are predicted to achieve better performance, e.g., reaction conditions that lead to higher yields. Moreover, machine learning models trained on published data can provide initial guesses for these optimizations, and even design full synthetic pathways for the synthesis of novel molecules.
Kerry Gilmore, Assistant Professor, Department of Chemistry, University of Connecticut
- Facilitating synthesis through standardization and predictive algorithms
- Abstract: Reaction and process development is traditionally a trial-and-error process. As such it is vulnerable to human error/biases and limited by the physical skills of the chemist. Automated flow chemistry platforms can significantly improve reproductivity, throughput, and decouple physical skills with reaction outcome. However, the efficiency of optimization and process development is still dependent on the chemist being able to navigate multi-dimensional space as well as observe, quantify, and extrapolate from the independent and interdependent factors influencing a reaction’s outcome. Machine learning algorithms can perform these tasks, allowing for more efficient and effective reaction optimization and development.
Grace Russell, Scientist, Snapdragon Chemistry Inc.
- Automated, reliable reaction sampling and analysis for data rich organic synthesis and auto optimization
- Abstract: Reliable reaction sampling is known to be challenging due to the wide range of reaction conditions such as temperature, pressure and the physical nature of the mixture which causes data collected from organic reactions to be inconsistent or inaccurate. These obstacles make developing a method for reaction analysis and optimization non-trivial. At Snapdragon Chemistry we have created a system for both traditional batch and continuous flow reaction sampling that is compatible with a wide range of reaction conditions allowing reliable sampling and optimization.
Panel Discussion Moderated by Malika Jeffries-El, Associate Professor, Chemistry, Boston University
12:20pm-1:20pm ET: Lunch & Student Poster Session (in Gather.Town)
1:20pm – 2:50pm ET: Session 2 – Machine Learning Innovation in Support of Chemistry and Materials Science
This session will focus on some of the fundamental machine learning techniques and algorithms that are used for applications in chemistry and material science. The speakers in this session all have worked on fundamental learning tools for working with structured data such as networks of molecules. This work includes auto-encoders for molecular conformations, learning on graphs and combinatorial structures, generative models for molecular graphs, among others. In the session, we hope to cover and discuss general tools that can be used in a wide variety of contexts.
Frank Noe, Professor, Department of Mathematics and Computer Science, Freie Universität Berlin
- Deep learning for molecular sciences
- Abstract: AI, and specifically deep ML methods have a profound impact on industry and information technology. But since recently AI methods are also changing the way we do science. In this talk I will present some of our recent efforts to build machine learning methods that attack fundamental problems in physical and chemical sciences: the sampling problem in physical many-body systems, and the solution of the quantum-chemical electronic Schrödinger equation. Key in making progress in these hard problems with ML is to interrogate the physical system about what the learning problem should be, and to encode physical structures, such as symmetries and conservation laws, into the ML model.
Risi Kondor, Associate Professor, Departments of Computer Science and Statistics, The University of Chicago
- Equivariant neural neural networks for physics and chemistry
- Abstract: Unlike many other “big data” domains, when using machine learning in physics or chemistry, we have physical laws that the learning algorithm must satisfy, such as invariance to translations, rotations, and other symmetries. The theory of group equivariant neural networks, built on representation theory, provides a general framework for incorporating such symmetries in a computationally efficient manner. I will discuss this general theory, highlight some recent practical successes and describe some future directions.
Bidisha Samanta, Research Engineer, Google
- NeVAE: A deep generative model for molecular graphs
- Abstract: Deep generative models have been praised for their ability to learn smooth latent representation of images, text, and audio, which can then be used to generate new, plausible data. However, current generative models are unable to work with molecular graphs due to their unique characteristics—their underlying structure is not Euclidean or grid-like, they remain isomorphic under permutation of the nodes labels, and they come with a different number of nodes and edges. In our work we propose a variational autoencoder for molecular graphs, whose encoder and decoder are specially designed to account for the above properties by means of several technical innovations. We further develop a gradient-based algorithm to optimize the decoder of our model so that it learns to generate molecules that maximize the value of a certain property of interest and, given a molecule of interest, it is able to optimize the spatial configuration of its atoms for greater stability.
Panel Discussion Moderated by Brian Kulis, Associate Professor. Electrical and Computer Engineering, Boston University
2:50pm-3:05pm ET: Break
3:05pm – 4:35pm ET: Session 3 – Machine Learning in Computational Chemical and Materials Science
Computational methods are able to resolve chemical-physical phenomena from the electronic structure through the atomic and continuum levels; however they require high performance computing and long compute times. Integrating machine learning and data science approaches with computational methods has the potential to allow more effective and efficient chemical and materials development. In this session the promises and challenge of machine learning in computational chemical and materials science will be discussed by leaders in the field and through panel discussion of where research needs to go.
Pratyush Tiwary, Assistant Professor, Department of Chemistry & Biochemistry, University of Maryland
- Can artificial intelligence help predict how drugs unbind, proteins fold and snowflakes form?
- Abstract: The brain’s ability to rapidly learn from complex data to make predictions about the future is crucial to success in life. This could in baseball, stock trading or swatting flies. Modern day artificial intelligence (AI) mimics this fidelity and has been widely successful. Here I will show AI can also be used in chemistry to understand, and predict, the behavior of complex molecules with millions of atoms. However AI has several shortcomings, which we overcome through old and new ideas in the field of statistical mechanics. I will talk about such methods, successes so far and challenges ahead.
Michele Ceriotti, Associate Professor, Institute of Materials, Swiss Federal Institute of Technology Lausanne
- Atomistic simulations in the age of machine learning
- Abstract: When modeling materials and molecules at the atomic scale, achieving a realistic level of complexity and making quantitative predictions are usually conflicting goals. Data-driven techniques have made great strides towards enabling simulations of materials in realistic conditions with uncompromising accuracy. In this talk I will summarize the core concepts that have driven the extraordinarily fast progress of the field, discuss some of the most promising modeling techniques that combine physics-inspired and data-driven paradigms, indicate the most pressing open challenges, and present several compelling examples ranging from water to semiconductors and from metals to molecular materials.
Yu-Shan Lin, Associate Professor, Chemistry, Tufts University
- Structure Prediction of Cyclic Peptides via Structural Ensembles Achieved by Molecular Dynamics and Machine Learning (StrEAMM)
- Abstract: Cyclic peptides represent a special class of compounds in the “beyond rule-of-five” chemical space, and they have captured the attention of chemists and the pharmaceutical industry owing to their unique properties for therapeutic development. In this talk, I will discuss how by using molecular dynamics simulation results as training datasets, we developed machine-learning models that can predict the structural ensembles of cyclic peptides, which tend to adopt multiple low-population conformations in solution. Once trained, our models can predict in <0.1 second the solution structural ensemble for a cyclic pentapeptide that one would normally obtain only after running days of explicit-solvent molecular dynamics simulations using computational methods and resources normally accessible only to expert computation labs. The resulting method, termed StrEAMM (structural ensembles achieved by molecular dynamics and machine learning), is capable of efficiently predicting complete structural ensembles for cyclic peptides without relying on additional molecular dynamics simulations, constituting a seven-order-of-magnitude improvement in speed while retaining the same accuracy as explicit-solvent simulations. In one example, our training set includes 705 semi-randomly generated sequences to cover all the Xaa-Yaa-Zaa amino-acid patterns, and the resulting StrEAMM model replaces the need to run explicit-solvent simulations to provide structural ensemble prediction for the remaining >150,000 cyclic peptides in the sequence space.
Panel Discussion Moderated by Qiang Cui, Professor, Chemistry, Boston University
4:35pm-4:45pm ET: Closing Remarks
Hariri Institute for Computing, Boston University
About the Organizers
Visited 2 times, 1 Visit today