An official website of the United States government
Abstract In the last several years, there has been a surge in the development of machine learning potential (MLP) models for describing molecular systems. We are interested in a particular area of this field — the training of system‐specific MLPs for reactive systems — with the goal of using these MLPs to accelerate free energy simulations of chemical and enzyme reactions. To help new members in our labs become familiar with the basic techniques, we have put together a self‐guided Colab tutorial (https://cc-ats.github.io/mlp_tutorial/), which we expect to be also useful to other young researchers in the community. Our tutorial begins with the introduction of simple feedforward neural network (FNN) and kernel‐based (using Gaussian process regression, GPR) models by fitting the two‐dimensional Müller‐Brown potential. Subsequently, two simple descriptors are presented for extracting features of molecular systems: symmetry functions (including the ANI variant) and embedding neural networks (such as DeepPot‐SE). Lastly, these features will be fed into FNN and GPR models to reproduce the energies and forces for the molecular configurations in a Claisen rearrangement reaction.
more »
« less
PyXtal_FF: a python library for automated force field generation
Abstract We present PyXtal_FF—a package based on Python programming language—for developing machine learning potentials (MLPs). The aim of PyXtal_FF is to promote the application of atomistic simulations through providing several choices of atom-centered descriptors and machine learning regressions in one platform. Based on the given choice of descriptors (including the atom-centered symmetry functions, embedded atom density, SO4 bispectrum, and smooth SO3 power spectrum), PyXtal_FF can train MLPs with either generalized linear regression or neural network models, by simultaneously minimizing the errors of energy/forces/stress tensors in comparison with the data fromab-initiosimulations. The trained MLP model from PyXtal_FF is interfaced with the Atomic Simulation Environment (ASE) package, which allows different types of light-weight simulations such as geometry optimization, molecular dynamics simulation, and physical properties prediction. Finally, we will illustrate the performance of PyXtal_FF by applying it to investigate several material systems, including the bulk SiO2, high entropy alloy NbMoTaW, and elemental Pt for general purposes. Full documentation of PyXtal_FF is available athttps://pyxtal-ff.readthedocs.io.
more »
« less
- PAR ID:
- 10208347
- Publisher / Repository:
- IOP Publishing
- Date Published:
- Journal Name:
- Machine Learning: Science and Technology
- Volume:
- 2
- Issue:
- 2
- ISSN:
- 2632-2153
- Page Range / eLocation ID:
- Article No. 027001
- Format(s):
- Medium: X
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
Abstract As the next generation of large galaxy surveys come online, it is becoming increasingly important to develop and understand the machine-learning tools that analyze big astronomical data. Neural networks are powerful and capable of probing deep patterns in data, but they must be trained carefully on large and representative data sets. We present a new “hump” of the Cosmology and Astrophysics with MachinE Learning Simulations (CAMELS) project: CAMELS-SAM, encompassing one thousand dark-matter-only simulations of (100h−1cMpc)3with different cosmological parameters (Ωmandσ8) and run through the Santa Cruz semi-analytic model for galaxy formation over a broad range of astrophysical parameters. As a proof of concept for the power of this vast suite of simulated galaxies in a large volume and broad parameter space, we probe the power of simple clustering summary statistics to marginalize over astrophysics and constrain cosmology using neural networks. We use the two-point correlation, count-in-cells, and void probability functions, and we probe nonlinear and linear scales across 0.68 <R<27h−1cMpc. We find our neural networks can both marginalize over the uncertainties in astrophysics to constrain cosmology to 3%–8% error across various types of galaxy selections, while simultaneously learning about the SC-SAM astrophysical parameters. This work encompasses vital first steps toward creating algorithms able to marginalize over the uncertainties in our galaxy formation models and measure the underlying cosmology of our Universe. CAMELS-SAM has been publicly released alongside the rest of CAMELS, and it offers great potential to many applications of machine learning in astrophysics:https://camels-sam.readthedocs.io.more » « less
-
Abstract BackgroundComputational cell type deconvolution enables the estimation of cell type abundance from bulk tissues and is important for understanding tissue microenviroment, especially in tumor tissues. With rapid development of deconvolution methods, many benchmarking studies have been published aiming for a comprehensive evaluation for these methods. Benchmarking studies rely on cell-type resolved single-cell RNA-seq data to create simulated pseudobulk datasets by adding individual cells-types in controlled proportions. ResultsIn our work, we show that the standard application of this approach, which uses randomly selected single cells, regardless of the intrinsic difference between them, generates synthetic bulk expression values that lack appropriate biological variance. We demonstrate why and how the current bulk simulation pipeline with random cells is unrealistic and propose a heterogeneous simulation strategy as a solution. The heterogeneously simulated bulk samples match up with the variance observed in real bulk datasets and therefore provide concrete benefits for benchmarking in several ways. We demonstrate that conceptual classes of deconvolution methods differ dramatically in their robustness to heterogeneity with reference-free methods performing particularly poorly. For regression-based methods, the heterogeneous simulation provides an explicit framework to disentangle the contributions of reference construction and regression methods to performance. Finally, we perform an extensive benchmark of diverse methods across eight different datasets and find BayesPrism and a hybrid MuSiC/CIBERSORTx approach to be the top performers. ConclusionsOur heterogeneous bulk simulation method and the entire benchmarking framework is implemented in a user friendly packagehttps://github.com/humengying0907/deconvBenchmarkingandhttps://doi.org/10.5281/zenodo.8206516, enabling further developments in deconvolution methods.more » « less
-
Abstract We present an expansion of FLEET, a machine-learning algorithm optimized to select transients that are most likely tidal disruption events (TDEs). FLEET is based on a random forest algorithm trained on both the light curves and host galaxy information of 4779 spectroscopically classified transients. We find that for transients with a probability of being a TDE,P(TDE) > 0.5, we can successfully recover TDEs with ≈40% completeness and ≈30% purity when using their first 20 days of photometry or a similar completeness and ≈50% purity when including 40 days of photometry, an improvement of almost 2 orders of magnitude compared to random selection. Alternatively, we can recover TDEs with a maximum purity of ≈80% and a completeness of ≈30% when considering only transients withP(TDE) > 0.8. We explore the use of FLEET for future time-domain surveys such as the Legacy Survey of Space and Time on the Vera C. Rubin Observatory (Rubin) and the Nancy Grace Roman Space Telescope (Roman). We estimate that ∼104well-observed TDEs could be discovered every year by Rubin and ∼200 TDEs by Roman. Finally, we run FLEET on the TDEs from our Rubin survey simulation and find that we can recover ∼30% of them at redshiftz< 0.5 withP(TDE) > 0.5, or ∼3000 TDEs yr–1that FLEET could uncover from the Rubin stream. We have demonstrated that we will be able to run FLEET on Rubin photometry as soon as this survey begins. FLEET is provided as an open source package on GitHub: https://github.com/gmzsebastian/FLEET.more » « less
-
Abstract We compared the performance of DREAM3D simulations in reproducing the long‐term radiation belt dynamics observed by Van Allen Probes over the entire year of 2017 with various boundary conditions (BCs) and model inputs. Specifically, we investigated the effects of three different outer boundary conditions, two different low‐energy boundary conditions for seed electrons, four different radial diffusion (RD) coefficients (DLL), four hiss wave models, and two chorus wave models from the literature. Using the outer boundary condition driven by GOES data, our benchmark simulation generally well reproduces the observed radiation belt dynamics insideL* = 6, with a better model performance at lowerμthan higherμ, whereμis the first adiabatic invariant. By varying the boundary conditions and inputs, we find that: (a) The data‐driven outer boundary condition is critical to the model performance, while adding in the data‐driven seed population doesn't further improve the performance. (b) The model shows comparable performance withDLLfrom Brautigam and Albert (2000,https://doi.org/10.1029/1999ja900344), Ozeke et al. (2014,https://doi.org/10.1002/2013ja019204), and Liu et al. (2016,https://doi.org/10.1002/2015gl067398), while withDLLfrom Ali et al. (2016,https://doi.org/10.1002/2016ja023002) the model shows less RD compared to data. (c) The model performance is similar with data‐based hiss models, but the results show faster loss is still needed inside the plasmasphere. (d) The model performs similarly with the two different chorus models, but better capturing the electron enhancement at higherμusing the Wang et al. (2019,https://doi.org/10.1029/2018ja026183) model due to its stronger wave power, since local heating for higher energy electrons is under‐reproduced in the current model.more » « less
