Machine Learning in Multiphysics and Multiscale Computing

Back to 2019 MSM Agenda

Session Description:  This session will focus on machine learning (ML) algorithms for multiphysics and multiscale problems. There is much excitement about these algorithms, especially deep learning, in various disciplines, but their role in multiscale modeling and computational scientific problem solving is less clear. From experimental data to model development, solution, validation and uncertainty quantification as well as algorithm explanation, there are emerging efforts to integrate ML in the armamentarium of the MSM community. Our goal will be to review on-going research, investigate opportunities and challenges, and have an open discussion on the role that ML algorithms can play in multiscale modeling.

Speaker Bios and Abstracts:

4-4:15:  Ellen Kuhl, Stanford - Machine learning in drug development 



Abstract: An undesirable side effect of drugs are cardiac arrhythmias, in particular a condition called torsades de pointes. Current paradigms for drug safety evaluation are costly, lengthy, and conservative, and impede efficient drug development. Here we combine multiscale experiment and simulation, highperformance computing, and machine learning to create an easy-to-use risk assessment diagram to quickly and reliable stratify the pro-arrhythmic potential of new and existing drugs, see Figure 1 in attached document. We capitalize on recent developments in machine learning and integrate information across ten orders of magnitude in space and time to provide a holistic picture of the effects of drugs, either individually or in combination with other drugs. We show, both experimentally and computationally, that drug-induced arrhythmias are dominated by the interplay of two currents with opposing effects: the rapid delayed rectifier potassium current and the L-type calcium current. Using Gaussian process classification, we create a classifier that stratifies safe and arrhythmic domains for any combinations of these two currents. We demonstrate that our classifier correctly identifies the risk categories of 23 common drugs, exclusively on the basis of their concentrations at 50% current block. Our new risk assessment diagram explains under which conditions blocking the L-type calcium current can delay or even entirely suppress arrhythmogenic events. Using machine learning in drug safety evaluation can provide a more accurate and comprehensive mechanistic assessment of the pro-arrhythmic potential of new drugs. Our study shapes the way towards establishing science-based criteria to accelerate drug development, design safer drugs, and reduce heart rhythm disorders. 

Bio: Ellen Kuhl is Professor of Mechanical Engineering, Bioengineering, and Cardiothoracic Surgery and the current Chair of the US National Committee on Biomechanics. Her area of professional expertise is living matter physics, the creation of theoretical and computational models to predict the acute and chronic response of living structures to environmental changes during development and disease progression. Her specific interest is the multiscale modeling of growth and remodeling, the study of how living matter adapts its form and function to changes in mechanical loading, and how this adaptation can be traced back to structural alterations on the cellular or molecular levels. Growth and remodeling might be induced naturally, e.g., through elevated pressure, stress, or strain, or interventionally, e.g., through prostheses, stents, tissue grafts, or stem cell injection. Combining theories of applied mathematics, biophysics, and continuum mechanics, her lab has specialized in predicting the evolution of form and function in living structures using patient-specific custom-designed finite element models.

4:15-4:30:  George Karniadakis, Brown - Physics-Informed Neural Networks



Abstract: We will present a new approach to develop a data-driven, learning-based framework for predicting outcomes of physical and biological systems systems and for discovering hidden bio-physics from noisy data. We will introduce a deep learning approach based on neural networks (NNs) and generative adversarial networks (GANs). Unlike other approaches that rely on big data, here we “learn” from small data by exploiting the information provided by the physical and biochemical laws, which are used to obtain informative priors or regularize the neural networks.. We will demonstrate the power of PINNs for inverse problems in bio-fluid mechanics, e.g., inferring the wall shear stress in brain aneurysms from dye visualizations only, and also in systems biology by inferring the dynamics of yeast glycolysis, which has become a standard test case for biochemical dynamics inference.

Bio: George Karniadakis’ current research interests are on machine learning for scientific computing, that is how to solve and discover new PDEs via deep learning, hence removing the tyranny of grids and using gappy data only. His broad research interests focus on stochastic multiscale mathematics and modeling of physical and biological systems.  Current thrusts include probabilistic numerics, stochastic simulation (in the context of uncertainty quantification and beyond), fractional PDEs, and multiscale modeling of complex systems (e.g., the brain). His new area is neurovascular coupling in the brain, i.e., bridging the gap between neuroscience and vascular mechanics. New experimental evidence suggests the intriguing possibility that by slightly modulating the brain blood flow one can control information processing.

I don't quite understand how PINNS can converge to the correct solution even though the problems are nonlinear and only have partial information to go on. Wouldn't there by multiple solutions in such a system?
Herbert M Sauro

4:30-4:45:  Bill Cannon, Pacific Northwest National Laboratory - Using Principles of Self-Organization to Model Metabolism: Prediction of Rate Constants, Concentrations and Enzymatic Regulation


4:45-5:00 Danny Bluestein, Stony Brook University - Machine Learning in Multiscale Modeling of Blood Flow and Platelet Mediated Thrombosis


Abstract: We developed MSM approach by incorporating coarse-grained molecular dynamics (CGMD) and dissipative particle dynamics (DPD) to describe platelet mechanotransduction induced by blood flow in cardiovascular pathologies which may initiate thrombosis. While MSM sufficiently describes details at disparate spatial scales, no effective algorithm for adapting temporal scales to these diverse spatial scales exists. We propose a novel state-driven adaptive time-stepping (ATS) algorithm6,7 that adapts time stepsizes to the underlying biophysical phenomena: mesoscale DPD blood flow is simulated with timescale and microscale CGMD platelet is modeled with to timescales. A ML-based framework trains to adapt the time stepsizes. Particle positions and momenta are inputs, and phases are described by the most significant attributes of states from inputs in first two layers- categorized by a neural network and labeled by a two components vector: time stepsize and state examination frequency. The simulation proceeds with a new time stepsize in steps. The ATS algorithm adjusts time stepsizes at its conclusion. The ATS algorithm was compared with traditional single time-stepping (STS) algorithm with relative errors along time of system kinetic energy, and distance between center of mass of two platelets. The final states of aggregation in both algorithms are consistent with each other. Computing times using ATS for different simulations phases were cut by 20~75%. Conceptually, ATS ML corresponds to coarse-graining in time.

Bio: Dr. Bluestein's work tackles the dynamics of flow and cellular transport in blood recirculating devices and the diseased cardiovascular system. Non-physiological flow fields that arise as a result of cardiovascular diseases and prosthetic devices have a complex interaction with the blood vessels and the blood itself. Over the last decade, evidence has accumulated indicating that local flow induced mechanical forces alter the molecular mechanism of the formed elements of blood, and have a major effect on blood clotting. The end result can be thrombus formation that can occlude arteries, or handicap the functionality of implanted devices. The long term goal of Dr. Bluestein's work is to elucidate flow-induced pathologies in blood recirculating devices and the cardiovascular system in order to advance our understanding of the different aspects of flows in biological systems. This research involves the use of sophisticated non-invasive techniques, such as numerical modeling of devices and pathologies, multiscale modeling of platelets, and innovative biochemical assaying techniques.

2017 MSM Session on New Methods

2018 MSM Special Session on Machine Learning

2018 MSM Breakout on New Methods

Interactive Discussion (please put you name before your comments):


To the Panel: how would you

Your name
Gary An

To the Panel: how would you see the lessons/methods learned/employed in the DeepMind game-playing AIs (AlphaGo, AlphaZeroGo and AlphaZero) being applied to multi-scale modeling?

Submitted by Anonymous (not verified) on Wed, 03/06/2019 - 17:02

Answer to Gary. The

Your name
Jacob Barhak

Answer to Gary. The difference is that a machine playing a game can be easily repeated many times on many virtual machines. In modeling biology and physics many times, the experiments cannot be repeated easily.

Submitted by Anonymous (not verified) on Wed, 03/06/2019 - 17:18

Our paper on applying the

Your name
Gary An

Our paper on applying the game-playing AI approach to identifying precision therapies for sepsis (applicable to any disease process represented mechanistically). Journal of Computational BIology DOI: 10.1089/cmb.2018.0168

Submitted by Anonymous (not verified) on Wed, 03/06/2019 - 17:28

Table sorting checkbox