Transfer Learning from Simulation to Experimental Data: NMR Chemical Shift Predictions

Abstract

An accurate prediction of chemical shifts (δ) to elucidate molecular structures has been a challenging problem. Recently, noble machine learning architectures achieve accurate prediction performance, but the difficulty of building a huge chemical database limits the applicability of machine learning approaches. In this work, we demonstrate that the prior knowledge gained from the simulation database is successfully transferred into the problem of predicting an experimentally measured δ. Although both simulation and experimental databases are vastly different in chemical perspectives, reliable accuracy for δ is achieved by additional training with randomly sampled small numbers of experimental data. Furthermore, the prior knowledge allows us to successfully train the model on the more focused chemical space that the experimental database sparsely covers. The proposed approach, the knowledge transfer from the simulation database, can be utilized to enhance the usability of the local experimental database.

Prediction of chemical shift in NMR: A review

“…The ability to learn to predict without assigning an observed shift to a given nucleus [47] is potentially powerful and could resolve numerous dataset limitations outlined in Section 2.3. Han and Choi [48] use transfer learning with a large simulation database to be able to train models with small numbers of experimental data. It achieves comparable results as other deep learning approaches with a much smaller dataset.…”

Systematic Investigation of Error Distribution in Machine Learning Algorithms Applied to the Quantum-Chemistry QM9 Data Set Using the Bias and Variance Decomposition

“…It would also be interesting to carry out bias–variance analysis with other ML algorithms, such as deep neural networks. Furthermore, it would be interesting to compare the bias–variance decomposition with uncertainty estimation using Monte Carlo dropout, as used in refs and .…”

“…The authors claim that the average model test error is an overly simplistic metric to assess the quality or failure of an ML model as the underlying choices involved in designing the model may not be adequate for the entire class of materials. Kwon et al 17 and Han and Choi 18 use a Monte Carlo drop out rule in message passing neural networks to estimate prediction uncertainty and design a regret classification rule for nuclear magnetic resonance chemical shift prediction. Pernot et al 19 discuss using different prediction uncertainty measures to evaluate the performance of models associated with a posteriori model calibration to predict solids' properties.…”

Increasing Efficiency of Nonadiabatic Molecular Dynamics by Hamiltonian Interpolation with Kernel Ridge Regression

“…Machine learning (ML) techniques have experienced rapid and extensive developments in recent decades, stimulated by growing computational power and large amounts of experimental and theoretical data accumulated in all fields, including science, technology, and daily life. Many efforts are devoted to applying ML to chemistry and materials science to predict relevant properties, − identify promising candidates for various applications, − and guide and design ongoing experiments. , With the help of ab initio-calculated data, ML is successfully practiced for predicting molecular properties. − Prediction of time-series data, such as forces on atoms in molecular dynamics (MD) simulation, can alleviate the burden of the computational cost of quantum mechanical calculations . The key to the success stems from the fact that ML models can quantitatively estimate the behavior in an unknown realm by learning the pattern from an existing data set.…”