Machine learning prediction errors better than DFT accuracy

Abstract

We investigate the impact of choosing regressors and molecular representations for the construction of fast machine learning (ML) models of thirteen electronic ground-state properties of organic molecules. The performance of each regressor/representation/property combination is assessed using learning curves which report out-of-sample errors as a function of training set size with up to $\sim$117k distinct molecules. Molecular structures and properties at hybrid density functional theory (DFT) level of theory used for training and testing come from the QM9 database [Ramakrishnan et al, {\em Scientific Data} {\bf 1} 140022 (2014)] and include dipole moment, polarizability, HOMO/LUMO energies and gap, electronic spatial extent, zero point vibrational energy, enthalpies and free energies of atomization, heat capacity and the highest fundamental vibrational frequency. Various representations from the literature have been studied (Coulomb matrix, bag of bonds, BAML and ECFP4, molecular graphs (MG)), as well as newly developed distribution based variants including histograms of distances (HD), and angles (HDA/MARAD), and dihedrals (HDAD). Regressors include linear models (Bayesian ridge regression (BR) and linear regression with elastic net regularization (EN)), random forest (RF), kernel ridge regression (KRR) and two types of neural net works, graph convolutions (GC) and gated graph networks (GG). We present numerical evidence that ML model predictions deviate from DFT less than DFT deviates from experiment for all properties. Furthermore, our out-of-sample prediction errors with respect to hybrid DFT reference are on par with, or close to, chemical accuracy. Our findings suggest that ML models could be more accurate than hybrid DFT if explicitly electron correlated quantum (or experimental) data was available.

References in corpus (8)

• Neural Message Passing for Quantum Chemistry
• ANI-1: An extensible neural network potential with DFT accuracy at force field computational cost
• Machine Learning Unifies the Modelling of Materials and Molecules
• Unified Representation of Molecules and Crystals for Machine Learning
• Understanding molecular representations in machine learning: The role of uniqueness and target similarity
• Accelerated materials property predictions and design using motif-based fingerprints
• Constant Size Molecular Descriptors For Use With Machine Learning
• Localized Coulomb Descriptors for the Gaussian Approximation Potential

Cited by in corpus (138)

• Machine learning and the physical sciences
• SchNet - a deep learning architecture for molecules and materials
• Graph Networks as a Universal Machine Learning Framework for Molecules and Crystals
• Recent Advances and Applications of Deep Learning Methods in Materials Science
• PhysNet: A Neural Network for Predicting Energies, Forces, Dipole Moments and Partial Charges
• Combining Machine Learning and Computational Chemistry for Predictive Insights Into Chemical Systems
• DScribe: Library of Descriptors for Machine Learning in Materials Science
• Machine Learning Unifies the Modelling of Materials and Molecules
• Atomistic Line Graph Neural Network for Improved Materials Property Predictions
• Deep learning for molecular design - a review of the state of the art
• Towards Exact Molecular Dynamics Simulations with Machine-Learned Force Fields
• Tensor field networks: Rotation- and translation-equivariant neural networks for 3D point clouds
• Physics-inspired structural representations for molecules and materials
• Machine learning a general purpose interatomic potential for silicon
• SchNet: A continuous-filter convolutional neural network for modeling quantum interactions
• SchNetPack: A Deep Learning Toolbox For Atomistic Systems
• ANI-1: A data set of 20M off-equilibrium DFT calculations for organic molecules
• FCHL revisited: faster and more accurate quantum machine learning
• Alchemical and structural distribution based representation for improved QML
• Resolving transition metal chemical space: feature selection for machine learning and structure-property relationships
• Hierarchical modeling of molecular energies using a deep neural network
• Automatic Selection of Atomic Fingerprints and Reference Configurations for Machine-Learning Potentials
• WACSF - Weighted Atom-Centered Symmetry Functions as Descriptors in Machine Learning Potentials
• Unified Representation of Molecules and Crystals for Machine Learning
• Autonomous discovery in the chemical sciences part II: Outlook
• Applying machine learning techniques to predict the properties of energetic materials
• sGDML: Constructing Accurate and Data Efficient Molecular Force Fields Using Machine Learning
• Accurate molecular polarizabilities with coupled-cluster theory and machine learning
• Atom2Vec: learning atoms for materials discovery
• Deep learning and the Schrödinger equation
• Non-covalent interactions across organic and biological subsets of chemical space: Physics-based potentials parametrized from machine learning
• Machine learning of molecular properties: locality and active learning
• Representations of molecules and materials for interpolation of quantum-mechanical simulations via machine learning
• Feature Optimization for Atomistic Machine Learning Yields A Data-Driven Construction of the Periodic Table of the Elements
• Neural Network Potentials for Chemistry: Concepts, Applications and Prospects
• The Modern Mathematics of Deep Learning
• Deep Learning and Knowledge-Based Methods for Computer Aided Molecular Design -- Toward a Unified Approach: State-of-the-Art and Future Directions
• A Deep Dive into Machine Learning Density Functional Theory for Materials Science and Chemistry
• Message-passing neural networks for high-throughput polymer screening
• Operators in Machine Learning: Response Properties in Chemical Space
• Molecular Force Fields with Gradient-Domain Machine Learning: Construction and Application to Dynamics of Small Molecules with Coupled Cluster Forces
• Learning neural network potentials from experimental data via Differentiable Trajectory Reweighting
• Accelerating CALYPSO Structure Prediction by Data-driven Learning of Potential Energy Surface
• Learning the exchange-correlation functional from nature with fully differentiable density functional theory
• Graph convolutional neural networks as "general-purpose" property predictors: the universality and limits of applicability
• Image-based deep learning for classification of noise transients in gravitational wave detectors
• SchNetPack 2.0: A neural network toolbox for atomistic machine learning
• Error-Controlled Exploration of Chemical Reaction Networks with Gaussian Processes
• Quantum deep field: data-driven wave function, electron density generation, and atomization energy prediction and extrapolation with machine learning
• Chemical diversity in molecular orbital energy predictions with kernel ridge regression
• Deeply learning molecular structure-property relationships using attention- and gate-augmented graph convolutional network
• Recursive evaluation and iterative contraction of $N$-body equivariant features
• Neural Message Passing with Edge Updates for Predicting Properties of Molecules and Materials
• The Evolution of Machine Learning Potentials for Molecules, Reactions and Materials
• Probing out-of-distribution generalization in machine learning for materials
• Hierarchical, rotation-equivariant neural networks to select structural models of protein complexes
• Detection of Topological Materials with Machine Learning
• Machine Learning of Free Energies in Chemical Compound Space Using Ensemble Representations: Reaching Experimental Uncertainty for Solvation
• Putting Density Functional Theory to the Test in Machine-Learning-Accelerated Materials Discovery
• High-Dimensional Potential Energy Surfaces for Molecular Simulations
• Simulations meet Machine Learning in Structural Biology
• Molecular Force Fields with Gradient-Domain Machine Learning (GDML): Comparison and Synergies with Classical Force Fields
• Molecular Mechanics-Driven Graph Neural Network with Multiplex Graph for Molecular Structures
• Challenges for Machine Learning Force Fields in Reproducing Potential Energy Surfaces of Flexible Molecules
• Accurate machine learning force fields via experimental and simulation data fusion
• Motif-based Graph Self-Supervised Learning for Molecular Property Prediction
• Prediction of Atomization Energy Using Graph Kernel and Active Learning
• Difficulty in chirality recognition for Transformer architectures learning chemical structures from string
• Machine learning the computational cost of quantum chemistry
• Atomistic structure learning
• A New Kind of Atlas of Zeolite Building Blocks
• ML Models of Vibrating H${}_2$CO: Comparing Reproducing Kernels, FCHL and PhysNet
• Improved machine learning algorithm for predicting ground state properties
• Predicting the phase diagram of titanium dioxide with random search and pattern recognition
• Deep Neural Network Computes Electron Densities and Energies of a Large Set of Organic Molecules Faster than Density Functional Theory (DFT)
• The long road to calibrated prediction uncertainty in computational chemistry
• Impact of non-normal error distributions on the benchmarking and ranking of Quantum Machine Learning models
• Lightweight and Effective Tensor Sensitivity for Atomistic Neural Networks
• Generating equilibrium molecules with deep neural networks
• Impact of the characteristics of quantum chemical databases on machine learning predictions of tautomerization energies
• 3DMolNet: A Generative Network for Molecular Structures
• A Universal Framework for Accurate and Efficient Geometric Deep Learning of Molecular Systems
• Enhancing the Accuracy of Density Functional Tight Binding Models Through ChIMES Many-body Interaction Potentials
• Controlled exploration of chemical space by machine learning of coarse-grained representations
• Kernel based quantum machine learning at record rate : Many-body distribution functionals as compact representations
• 3DReact: Geometric deep learning for chemical reactions
• Optimal radial basis for density-based atomic representations
• Leaving the Valley: Charting the Energy Landscape of Metal/Organic Interfaces via Machine Learning
• Effects of perturbation order and basis set on alchemical predictions
• Reducing Training Data Needs with Minimal Multilevel Machine Learning (M3L)
• Ab initio machine learning of phase space averages
• Machine Learning for compositional disorder: A Comparison Between Different Descriptors and Machine Learning Frameworks
• Transferable and Robust Machine Learning Model for Predicting Stability of Si Anodes for Multivalent Cation Batteries
• Wasserstein metric for improved QML with adjacency matrix representations
• Learning the energy curvature versus particle number in approximate density functionals
• Alchemical geometry relaxation
• Machine-learning of atomic-scale properties based on physical principles
• Heterogeneous Molecular Graph Neural Networks for Predicting Molecule Properties
• Deep learning and high harmonic generation
• Supervised learning of few dirty bosons with variable particle number
• Construction of Machine Learned Force Fields with Quantum Chemical Accuracy: Applications and Chemical Insights
• Accelerated Bayesian Inference for Molecular Simulations using Local Gaussian Process Surrogate Models
• Efficient prediction of 3D electron densities using machine learning
• Supervised learning of random quantum circuits via scalable neural networks
• Accurate Molecular Dynamics Enabled by Efficient Physically-Constrained Machine Learning Approaches
• Machine Learning for Screening Large Organic Molecules
• SPA${}^{\mathrm{H}}$M(a,b): encoding the density information from guess Hamiltonian in quantum machine learning representations
• Shadow molecular dynamics and atomic cluster expansions for flexible charge models
• Machine learning at the atomic-scale
• Compressing physical properties of atomic species for improving predictive chemistry
• Discovery of structure-property relations for molecules via hypothesis-driven active learning over the chemical space
• Machine Learning Framework for Modeling Exciton-Polaritons in Molecular Materials
• Machine learning frontier orbital energies of nanodiamonds
• Electronic Descriptors for Supervised Spectroscopic Predictions
• AI-driven Inverse Design System for Organic Molecules
• Molecular Hessian matrices from a machine learning random forest regression algorithm
• Deep Neural Networks Are Effective At Learning High-Dimensional Hilbert-Valued Functions From Limited Data
• Machine Learning S-Wave Scattering Phase Shifts Bypassing the Radial Schrödinger Equation
• Physical machine learning outperforms "human learning" in Quantum Chemistry
• Ab initio machine learning in chemical compound space
• Independent Vector Analysis for Data Fusion Prior to Molecular Property Prediction with Machine Learning
• Artificial Intelligence based Autonomous Molecular Design for Medical Therapeutic: A Perspective
• Improving Malware Detection Accuracy by Extracting Icon Information
• Uncertainty quantification of molecular property prediction using Bayesian neural network models
• Extending the definition of atomic basis sets to atoms with fractional nuclear charge
• Predicting Kovats Retention Indices Using Graph Neural Networks
• Learning to Make Chemical Predictions: the Interplay of Feature Representation, Data, and Machine Learning Algorithms
• Prediction of Atomization Energies of Au13+ Clusters through the Machine Learning Approach
• Efficient interpolation of molecular properties across chemical compound space with low-dimensional descriptors
• InteractionNet: Modeling and Explaining of Noncovalent Protein-Ligand Interactions with Noncovalent Graph Neural Network and Layer-Wise Relevance Propagation
• Improving accuracy of interatomic potentials: more physics or more data? A case study of silica
• Gap prediction in hybrid graphene - hexagonal boron nitride nanoflakes using artificial neural networks
• Combining DFT with ML to study size specific interactions between metal clusters and adsorbates
• Analysis of Atomistic Representations Using Weighted Skip-Connections
• Batched Sparse Matrix Multiplication for Accelerating Graph Convolutional Networks
• Accelerated Simulations of Molecular Systems through Learning of their Effective Dynamics
• Learning the gravitational force law and other analytic functions
• Bond type restricted radial distribution functions for accurate machine learning prediction of atomization energies