Table 2.

Summary of advances in applying ML to energy harvesting, storage and conversion

ML approach	Main research outcome
Photovoltaics
Bayesian optimization	By sampling just 1.8% of the compositional space, the perovskites identified showed a >17-fold stability improvement over the original methylammonium lead iodide (MAPbI3) without compromising conversion efficiency160
Random forest classifier	Approach reliability was verified by screening ten newly designed donor materials, with good consistency between model predictions and experimental outcomes161
ML regression algorithms	Six lead-free hybrid perovskites that are stable at room temperature and have suitable bandgaps for solar cells were successfully screened out of 5,158 candidates23
Gaussian process regression	The ML model was able to make bandgap predictions of elpasolite compounds with similar accuracy to that of high-cost computational calculations21
KRR	Starting from a set of 1.2 million features, two of them were identified as the most important factors that influence the bandgap of double perovskites22
Random forest regressor	Compared with the brute-force method, an acceleration factor of 700 was achieved with an experimentally validated new perovskite24
DNN-based classifier	The ML model could classify compounds ten times faster than human analysis with 90% accuracy, and four lead-free layered perovskites were realized experimentally162
CNN-based classifier and random forest regressor	CNN-based crystal recognition enabled autonomous characterization of the outcomes of the robotic experiments. The regressor predicted the optimal conditions for the synthesis of a new perovskite single crystal163
Bayesian optimization	Using an automated experimentation platform with a Bayesian optimization, a four-dimensional parameter space of organic photovoltaics blends was mapped and optimized for photostability164
Random forest regressor	Essential features were identified and used to screen the most promising thin low-dimensional perovskite capping layer, which then led to a many-fold increase in the stability of the state-of-the-art perovskite cell165
Random forest regressor	Major patterns regarding materials selection/device structure were captured, which could be used to predict perovskite solar cell efficiencies166
Genetic algorithm	Experimental samples processed under conditions suggested by the model showed substantial improvements in performance167
Batteries
DNNs, KRR and support vector machine	The model enabled a reduction in the amount of density functional theory calculations required to explore the chemical space. Up to 5,000 candidate materials for Na-ion and K-ion electrodes were identified45
Artificial neural network	The model demonstrated accurate estimation of the redox potentials of molecular electrode materials in Li-ion batteries, with contribution analysis confirming that electron affinity has the highest contribution to the redox potential168
Gaussian KRR and GBR	The method predicted the redox potentials well. The redox potentials could be explained by a small number of features, improving the interpretability of the results169
Logistic regression	The screening reduced the list of candidate materials from 12,831 down to 21 structures that showed promise as electrolytes36
Linear regression and support vector machine	The method transferred physical insights onto more generic descriptors, allowing the screening of billions of unknown compositions for Li-ion conductivity46
Logistic regression	The ML-guided search was 2.7 times more likely to identify fast Li ion conductors, with at least a 44-fold improvement of room-temperature Li ion conductivity47
Hierarchical and spectral clustering	Ab initio molecular dynamics simulations were used to validate the clustering in Li-containing compounds and identify top candidates for high ionic conduction, with 16 new Li-ion conductors discovered38
Artificial neural network	Predicted electrode specific resistances were found to agree well with simulated values170
Crystal graph CNN, KRR and GBR	The ML model was used to screen over 12,000 inorganic solids for their use as solid electrolytes. Four of these solid electrolytes could be used to suppress Li dendrite growth37
Model-free reinforcement learning	This method was used to explore trade-offs in the power–performance design space and converge to a better power management policy. Experimental results obtained with this technique exhibited a remarkable power reduction compared with the existing expert-based power management171
Bagged decision tree	The model led to a policy for battery usage optimization that substantially outperformed the leading algorithms. The policy was capable of improving and adapting as new data were collected over time172
Electrocatalysis
Random forest regressor and extra trees regressor	The framework was able to identify 131 intermetallic surfaces across 54 alloys as promising candidates for CO2 reduction. Specifically, a Cu–Al alloy catalyst was identified and experimentally verified to selectively convert CO2 into ethylene with record performance75,76
Neural networks	The model reduced the number of intermediate ab initio calculations needed to locate saddle points on the potential-energy surface using a nudged elastic band simulation69
Gaussian process regressor	The model predicted the most important reaction step that needed to be calculated with the computationally demanding electronic structure theory. Using this method, the most likely reaction mechanism for the reaction of syngas on Rd(111) was identified70
Neural networks	A neural network was able to screen for active sites across a random, disordered nanoparticle surface. The most likely active sites for CO2 conversion were identified for Au and Cu nanoparticle systems71,72

CNN, convolutional neural network; DNN, deep neural network; GBR, gradient boosting regression; KRR, kernel ridge regression; ML, machine learning.