Machine Learning for Accurate Force Calculations in Molecular Dynamics Simulations
read more
Citations
Machine Learning of Molecular Electronic Properties in Chemical Compound Space
Neural Network Potential Energy Surfaces for Small Molecules and Reactions
Fourier series of atomic radial distribution functions: A molecular fingerprint for machine learning models of quantum chemical properties
A review on machine learning algorithms for the ionic liquid chemical space
DeepPocket: Ligand Binding Site Detection and Segmentation using 3D Convolutional Neural Networks.
References
Adam: A Method for Stochastic Optimization
Development of the Colle-Salvetti correlation-energy formula into a functional of the electron density
CRC Handbook of Chemistry and Physics
Density-functional exchange-energy approximation with correct asymptotic behavior.
Multilayer feedforward networks are universal approximators
Related Papers (5)
Generalized neural-network representation of high-dimensional potential-energy surfaces.
Frequently Asked Questions (17)
Q2. What have the authors stated for future works in "Machine learning for accurate force calculations in molecular dynamics simulations" ?
Di↵usion coe cient and viscosity calculations indicate that the new forces bias the simulation closer towards experimental data, indicating that these properties calculated from a long DFT simulation would most likely be closer to experimental values as compared to a purely classical force field. Future work to modify the feature vector to extend this method to complex multicomponent systems is in progress. Finally, the time comparison data further emphasizes the e ciency of the model to run long and multiple replicate simulations which are vital in the calculation of thermodynamic and kinetic properties.
Q3. What is the common method used to train ML models?
ML models are often trained on reference data obtained from QMbased methods, such as, density functional theory (DFT) within the Kohn-Sham formalism.
Q4. Why are these regions excluded from the fitting procedure?
Due to the wellknown,84 large oscillations at very short times and noise at very long times, these regions are excluded from the fitting procedure.
Q5. What is the common method used to predict electronic properties of molecules?
The bag-of-bonds model was used to predict accurate electronic properties of molecules, such as, their polarizability and molecular frontier orbital energies.
Q6. Why was DFT chosen as the ab initio method to calculate forces?
Calculating forces for each frameDFT was chosen as the ab initio method to calculate forces because of its speed of calculations, so that a dataset can be generated in a reasonable amount of time.
Q7. What is the successful approach to predicting bulk phase systems?
One of the most successful approaches relies on quantum mechanical (QM) calculations on gas-phase (sometimes considering the implicit solvent model) clusters to parameterize a model meant for bulk phase simulations.
Q8. How many ANNs have been used to discover new inorganic materials?
16 ANNs have also been used along with genetic algorithm (GA) optimization to discover unconventional spin-crossover complexes, which emphasizes their power for discovering new inorganic materials.
Q9. What is the effect of random sampling on the atoms?
random sampling makes it very hard to place all points at some distance apart from each other, which is essential in their case to avoid two atoms being too close together resulting in non-physical contacts.
Q10. What is the role of ML in predicting atomic forces?
As the area of development of ML-FFs for MD simulations is expanding towards assessing and improving the accuracy and transferability of the model, learning and predicting atomic forces have been receiving notable successes.
Q11. What is the general shape of the plots at both pressures?
The general shape of the plots at both pressures indicate that the model generated trajectories follow the trend of the classical trajectories, due to the Fclassical contribution in the force model.
Q12. Why is it not feasible to calculate diusion coe cients using a?
Since di↵usion coe cient calculations require longer time scale trajectories (especially for gaseous states, owing to the high mean free path),77 it is not feasible to calculate this property using a DFT trajectory.
Q13. How have multiple electronic, ground, and excited-state properties been predicted simultaneously?
19 Multiple electronic, ground, and excited-state properties have also been predicted simultaneously using Coulomb matrices in conjunction with deep multi-task artificial neural networks.
Q14. How can the authors test the eectiveness of the trained model in real world scenarios?
The e↵ectiveness of the trained model in real world scenarios can be better understood by calculating physical properties using trajectories generated by using the neural network in addition to the ability to predict the target force data.
Q15. What is the need to construct a multiscale model to capture accurate dynamics of chemical processes?
The need to construct a multiscale model (considering electronic, nuclear dynamics5–7 and their coupling to slower, cooperative motions of the system) to capture accurate dynamics of chemical processes cannot be overstated.
Q16. What are the effects of fitted potentials on the calculation of physical properties?
The e↵ects of such fitted potentials on the calculation of physical properties obtained from their trajectories, at di↵erent physical conditions, such as, temperature and pressure, need to be studied to further reinforce on their future applications.
Q17. How many argon atoms were in the -NetFF model?
The code for the-NetFF model is freely available from https://github.com/devalab/delNetFF.The primary system used for generating the data for ML contained 96 argon atoms in a cubic box of size 16.7 x 16.7 x 16.7 Å3.