skip to main content

Title: INDEEDopt: a deep learning-based ReaxFF parameterization framework

Empirical interatomic potentials require optimization of force field parameters to tune interatomic interactions to mimic ones obtained by quantum chemistry-based methods. The optimization of the parameters is complex and requires the development of new techniques. Here, we propose an INitial-DEsign Enhanced Deep learning-based OPTimization (INDEEDopt) framework to accelerate and improve the quality of the ReaxFF parameterization. The procedure starts with a Latin Hypercube Design (LHD) algorithm that is used to explore the parameter landscape extensively. The LHD passes the information about explored regions to a deep learning model, which finds the minimum discrepancy regions and eliminates unfeasible regions, and constructs a more comprehensive understanding of physically meaningful parameter space. We demonstrate the procedure here for the parameterization of a nickel–chromium binary force field and a tungsten–sulfide–carbon–oxygen–hydrogen quinary force field. We show that INDEEDopt produces improved accuracies in shorter development time compared to the conventional optimization method.

; ; ; ; ; ;
Award ID(s):
Publication Date:
Journal Name:
npj Computational Materials
Nature Publishing Group
Sponsoring Org:
National Science Foundation
More Like this
  1. The development of reliable, yet computationally efficient interatomic forcefields is key to facilitate the modeling of glasses. However, the parameterization of novel forcefields is challenging as the high number of parameters renders traditional optimization methods inefficient or subject to bias. Here, we present a new parameterization method based on machine learning, which combines ab initio molecular dynamics simulations and Bayesian optimization. By taking the example of glassy silica, we show that our method yields a new interatomic forcefield that offers an unprecedented agreement with ab initio simulations. This method offers a new route to efficiently parameterize new interatomic forcefields for disordered solids in a non-biased fashion.
  2. Abstract Background

    Optimization of DNA and protein sequences based on Machine Learning models is becoming a powerful tool for molecular design. Activation maximization offers a simple design strategy for differentiable models: one-hot coded sequences are first approximated by a continuous representation, which is then iteratively optimized with respect to the predictor oracle by gradient ascent. While elegant, the current version of the method suffers from vanishing gradients and may cause predictor pathologies leading to poor convergence.


    Here, we introduce Fast SeqProp, an improved activation maximization method that combines straight-through approximation with normalization across the parameters of the input sequence distribution. Fast SeqProp overcomes bottlenecks in earlier methods arising from input parameters becoming skewed during optimization. Compared to prior methods, Fast SeqProp results in up to 100-fold faster convergence while also finding improved fitness optima for many applications. We demonstrate Fast SeqProp’s capabilities by designing DNA and protein sequences for six deep learning predictors, including a protein structure predictor.


    Fast SeqProp offers a reliable and efficient method for general-purpose sequence optimization through a differentiable fitness predictor. As demonstrated on a variety of deep learning models, the method is widely applicable, and can incorporate various regularization techniques to maintain confidence in the sequencemore »designs. As a design tool, Fast SeqProp may aid in the development of novel molecules, drug therapies and vaccines.

    « less
  3. Abstract

    Machine learning interatomic force fields are promising for combining high computational efficiency and accuracy in modeling quantum interactions and simulating atomistic dynamics. Active learning methods have been recently developed to train force fields efficiently and automatically. Among them, Bayesian active learning utilizes principled uncertainty quantification to make data acquisition decisions. In this work, we present a general Bayesian active learning workflow, where the force field is constructed from a sparse Gaussian process regression model based on atomic cluster expansion descriptors. To circumvent the high computational cost of the sparse Gaussian process uncertainty calculation, we formulate a high-performance approximate mapping of the uncertainty and demonstrate a speedup of several orders of magnitude. We demonstrate the autonomous active learning workflow by training a Bayesian force field model for silicon carbide (SiC) polymorphs in only a few days of computer time and show that pressure-induced phase transformations are accurately captured. The resulting model exhibits close agreement with both ab initio calculations and experimental measurements, and outperforms existing empirical models on vibrational and thermal properties. The active learning workflow readily generalizes to a wide range of material systems and accelerates their computational understanding.

  4. Abstract

    The earlier integration of validated Lennard–Jones (LJ) potentials for 8 fcc metals into materials and biomolecular force fields has advanced multiple research fields, for example, metal–electrolyte interfaces, recognition of biomolecules, colloidal assembly of metal nanostructures, alloys, and catalysis. Here we introduce 12-6 and 9-6 LJ parameters for classical all-atom simulations of 10 further fcc metals (Ac, Ca (α), Ce (γ), Es (β), Fe (γ), Ir, Rh, Sr (α), Th (α), Yb (β)) and stainless steel. The parameters reproduce lattice constants, surface energies, water interfacial energies, and interactions with (bio)organic molecules in 0.1 to 5% agreement with experiment, as well as qualitative mechanical properties under standard conditions. Deviations are reduced up to a factor of one hundred in comparison to earlier Lennard–Jones parameters, embedded atom models, and density functional theory. We also explain a quantitative correlation between atomization energies from experiments and surface energies that supports parameter development. The models are computationally very efficient and applicable to an exponential space of alloys. Compatibility with a wide range of force fields such as the Interface force field (IFF), AMBER, CHARMM, COMPASS, CVFF, DREIDING, OPLS-AA, and PCFF enables reliable simulations of nanostructures up to millions of atoms and microsecond time scales. User-friendlymore »model building and input generation are available in the CHARMM-GUI Nanomaterial Modeler. As a limitation, deviations in mechanical properties vary and are comparable to DFT methods. We discuss the incorporation of reactivity and features of the electronic structure to expand the range of applications and further increase the accuracy.

    « less
  5. Abstract

    Strong surface winds under extratropical cyclones exert intense surface stresses on the ocean that lead to upper-ocean mixing, intensified heat fluxes, and the generation of waves, that, over time, lead to swell waves (longer than 10-s period) that travel long distances. Because low-frequency swell propagates faster than high-frequency swell, the frequency dependence of swell arrival times at a measurement site can be used to infer the distance and time that the wave has traveled from its generation site. This study presents a methodology that employs spectrograms of ocean swell from point observations on the Ross Ice Shelf (RIS) to verify the position of high wind speed areas over the Southern Ocean, and therefore of extratropical cyclones. The focus here is on the implementation and robustness of the methodology in order to lay the groundwork for future broad application to verify Southern Ocean storm positions from atmospheric reanalysis data. The method developed here combines linear swell dispersion with a parametric wave model to construct a time- and frequency-dependent model of the dispersed swell arrivals in spectrograms of seismic observations on the RIS. A two-step optimization procedure (deep learning) of gradient descent and Monte Carlo sampling allows detailed estimates of themore »parameter distributions, with robust estimates of swell origins. Median uncertainties of swell source locations are 110 km in radial distance and 2 h in time. The uncertainties are derived from RIS observations and the model, rather than an assumed distribution. This method is an example of supervised machine learning informed by physical first principles in order to facilitate parameter interpretation in the physical domain.

    « less