skip to main content


Title: Splash in a flash: sharpness-aware minimization for efficient liquid splash simulation
We present sharpness-aware minimization (SAM) for fluid dynamics which can efficiently learn the plausible dynamics of liquid splashes. Due to its ability to achieve robust and generalizing solutions, SAM efficiently converges to a parameter set that predicts plausible dynamics of elusive liquid splashes. Our training scheme requires 6 times smaller number of epochs to converge and, 4 times shorter wall-clock time. Our result shows that sharpness of loss function has a close connection to the plausibility of fluid dynamics and suggests further applicability of SAM to machine learning based fluid simulation.  more » « less
Award ID(s):
2118061
NSF-PAR ID:
10343118
Author(s) / Creator(s):
Date Published:
Journal Name:
Annual Conference of the European Association for Computer Graphics, Eurographics
Page Range / eLocation ID:
1003
Format(s):
Medium: X
Sponsoring Org:
National Science Foundation
More Like this
  1. Deep neural networks are often overparameterized and may not easily achieve model generalization. Adversarial training has shown effectiveness in improving generalization by regularizing the change of loss on top of adversarially chosen perturbations. The recently proposed sharpness-aware minimization (SAM) algorithm conducts adversarial weight perturbation, encouraging the model to converge to a flat minima. SAM finds a common adversarial weight perturbation per-batch. Although per-instance adversarial weight perturbations are stronger adversaries and can potentially lead to better generalization performance, their computational cost is very high and thus it is impossible to use per-instance perturbations efficiently in SAM. In this paper, we tackle this efficiency bottleneck and propose sharpness-aware minimization with dynamic reweighting (delta-SAM). Our theoretical analysis motivates that it is possible to approach the stronger, per-instance adversarial weight perturbations using reweighted per-batch weight perturbations. delta-SAM dynamically reweights perturbation within each batch according to the theoretically principled weighting factors, serving as a good approximation to per-instance perturbation. Experiments on various natural language understanding tasks demonstrate the effectiveness of delta-SAM. 
    more » « less
  2. Neural-network quantum molecular dynamics (NNQMD) simulations based on machine learning are revolutionizing atomistic simulations of materials by providing quantum-mechanical accuracy but orders-of-magnitude faster, illustrated by ACM Gordon Bell prize (2020) and finalist (2021). State-of-the-art (SOTA) NNQMD model founded on group theory featuring rotational equivari- ance and local descriptors has provided much higher accuracy and speed than those models, thus named Allegro (meaning fast). On massively parallel super- computers, however, it suffers a fidelity-scaling problem, where growing number of unphysical predictions of interatomic forces prohibits simulations involving larger numbers of atoms for longer times. Here, we solve this problem by com- bining the Allegro model with sharpness aware minimization (SAM) for enhanc- ing the robustness of model through improved smoothness of the loss landscape. The resulting Allegro-Legato (meaning fast and “smooth”) model was shown to elongate the time-to-failure tfailure, without sacrificing computational speed or accuracy. Specifically, Allegro-Legato exhibits much weaker dependence of time- to-failure on the problem size, t_failure = N^−0.14 (N is the number of atoms) compared to the SOTA Allegro model (t_failure ∝ N^−0.29), i.e., systematically delayed time-to-failure, thus allowing much larger and longer NNQMD simulations without failure. The model also exhibits excellent computational scalabil- ity and GPU acceleration on the Polaris supercomputer at Argonne Leadership Computing Facility. Such scalable, accurate, fast and robust NNQMD models will likely find broad applications in NNQMD simulations on emerging exaflop/s computers, with a specific example of accounting for nuclear quantum effects in the dynamics of ammonia to lay a foundation of the green ammonia technology for sustainability. 
    more » « less
  3. Iron (Fe) is ubiquitous in nature and found as Fe II or Fe III in minerals or as dissolved ions Fe 2+ or Fe 3+ in aqueous systems. The interactions of soluble Fe have important implications for fresh water and marine biogeochemical cycles, which have impacts on global terrestrial and atmospheric environments. Upon dissolution of Fe III into natural aquatic systems, organic carboxylic acids efficiently chelate Fe III to form [Fe III –carboxylate] 2+ complexes that undergo a wide range of photochemistry-induced radical reactions. The chemical composition and photochemical transformations of these mixtures are largely unknown, making it challenging to estimate their environmental impact. To investigate the photochemical process of Fe III –carboxylates at the molecular level, we conduct a comprehensive experimental study employing UV-visible spectroscopy, liquid chromatography coupled to photodiode array and high-resolution mass spectrometry detection, and oil immersion flow microscopy. In this study, aqueous solutions of Fe III –citrate were photolyzed under 365 nm light in an experimental setup with an apparent quantum yield of ( φ ) ∼0.02, followed by chemical analyses of reacted mixtures withdrawn at increment time intervals of the experiment. The apparent photochemical reaction kinetics of Fe 3+ –citrates (aq) were expressed as two generalized consecutive reactions of with the experimental rate constants of j 1 ∼ 0.12 min −1 and j 2 ∼ 0.05 min −1 , respectively. Molecular characterization results indicate that R and I consist of both water-soluble organic and Fe–organic species, while P compounds are a mixture of water-soluble and colloidal materials. The latter were identified as Fe–carbonaceous colloids formed at long photolysis times. The carbonaceous content of these colloids was identified as unsaturated organic species with low oxygen content and carbon with a reduced oxidation state, indicative of their plausible radical recombination mechanism under oxygen-deprived conditions typical for the extensively photolyzed mixtures. Based on the molecular characterization results, we discuss the comprehensive reaction mechanism of Fe III –citrate photochemistry and report on the formation of previously unexplored colloidal reaction products, which may contribute to atmospheric and terrestrial light-absorbing materials in aquatic environments. 
    more » « less
  4. Surface morphology, in addition to hydrophobic and electrostatic effects, can alter how proteins interact with solid surfaces. Understanding the heterogeneous dynamics of protein adsorption on surfaces with varying roughness is experimentally challenging. In this work, we use single-molecule fluorescence microscopy to study the adsorption of α-lactalbumin protein on the glass substrate covered with a self-assembled monolayer (SAM) with varying surface concentrations. Two distinct interaction mechanisms are observed: localized adsorption/desorption and continuous-time random walk (CTRW). We investigate the origin of these two populations by simultaneous single-molecule imaging of substrates with both bare glass and SAM-covered regions. SAM-covered areas of substrates are found to promote CTRW, whereas glass surfaces promote localized motion. Contact angle measurements and atomic force microscopy imaging show that increasing SAM concentration results in both increasing hydrophobicity and surface roughness. These properties lead to two opposing effects: increasing hydrophobicity promotes longer protein flights, but increasing surface roughness suppresses protein dynamics resulting in shorter residence times. Our studies suggest that controlling hydrophobicity and roughness, in addition to electrostatics, as independent parameters could provide a means to tune desirable or undesirable protein interactions with surfaces.

     
    more » « less
  5. The human tear film is a multilayer structure in which the dynamics are often strongly affected by a floating lipid layer. That layer has liquid crystalline characteristics and plays important roles in the health of the tear film. Previous models have treated the lipid layer as a Newtonian fluid in extensional flow. Motivated to develop a more realistic treatment, we present a model for the extensional flow of thin sheets of nematic liquid crystal. The rod-like molecules of these substances impart an elastic contribution to the rheology. We rescale a weakly elastic model due to Cummings et al. [“Extensional flow of nematic liquid crystal with an applied electric field,” Eur. J. Appl. Math. 25, 397–423 (2014).] to describe a lipid layer of moderate elasticity. The resulting system of two nonlinear partial differential equations for sheet thickness and axial velocity is fourth order in space, but still represents a significant reduction of the full system. We analyze solutions arising from several different boundary conditions, motivated by the underlying application, with particular focus on dynamics and underlying mechanisms under stretching. We solve the system numerically, via collocation with either finite difference or Chebyshev spectral discretization in space, together with implicit time stepping. At early times, depending on the initial film shape, pressure either aids or opposes extensional flow, which changes the free surface dynamics of the sheet and can lead to patterns reminiscent of those observed in tear films. We contrast this finding with the cases of weak elasticity and Newtonian flow, where the sheet retains the same qualitative shape throughout time. 
    more » « less