Machine Learning for a Sustainable Energy Future

Zhenpeng Yao; Yanwei Lum; Andrew Johnston; Luis Martin Mejia-Mendoza; Xin Zhou; Yonggang Wen; Alan Aspuru-Guzik; Edward H. Sargent; Zhi Wei Seh

Source-linked AI summary

Machine Learning for a Sustainable Energy Future

Zhenpeng Yao, Yanwei Lum, Andrew Johnston, Luis Martin Mejia-Mendoza, Xin Zhou, Yonggang Wen, Alan Aspuru-Guzik, Edward H. Sargent, Zhi Wei Seh

arXiv:2210.10391v1Oct 19, 2022cond-mat.mtrl-scics.LG

TL;DR

Accelerating the transition to sustainable energy requires advances across energy technologies, infrastructure, and policies. This review evaluates how machine learning can support energy materials discovery and management, finding evidence that it learns trends identified through decades of human research.

Problem
Accelerating renewable-energy growth requires advances in harvesting, storage, conversion, and management, while universal design principles for new battery materials remain undefined.
Method
This review surveys ML concepts, introduces Acc(X)eleration Performance Indicators, and evaluates ML applications across energy materials discovery, devices, and smart grids.
Results
The review finds conclusive evidence that ML can learn trends human researchers have identified through decades of energy research.
Takeaways & Limitations
ML is particularly well-suited to discovering new materials, although the field remains nascent and suitable methodologies are still emerging.
Takeaways & Limitations
Evaluating material stability remains expensive and slow because degradation is complex and existing microscopy and simulation techniques do not allow rapid assessment.

Abstract

from arXiv · show

Transitioning from fossil fuels to renewable energy sources is a critical global challenge; it demands advances at the levels of materials, devices, and systems for the efficient harvesting, storage, conversion, and management of renewable energy. Researchers globally have begun incorporating machine learning (ML) techniques with the aim of accelerating these advances. ML technologies leverage statistical trends in data to build models for prediction of material properties, generation of candidate structures, optimization of processes, among other uses; as a result, they can be incorporated into discovery and development pipelines to accelerate progress. Here we review recent advances in ML-driven energy research, outline current and future challenges, and describe what is required moving forward to best lever ML techniques. To start, we give an overview of key ML concepts. We then introduce a set of key performance indicators to help compare the benefits of different ML-accelerated workflows for energy research. We discuss and evaluate the latest advances in applying ML to the development of energy harvesting (photovoltaics), storage (batteries), conversion (electrocatalysis), and management (smart grids). Finally, we offer an outlook of potential research areas in the energy field that stand to further benefit from the application of ML.

Introduction … Generative materials design

The review frames ML as a way to accelerate sustainable-energy materials discovery, device development, and energy management, addressing lengthy discovery timelines and growing energy demand. It surveys ML concepts, acceleration metrics, closed-loop frameworks, and applications across energy technologies and smart grids.

Introduction: Fossil fuels supply ~80% of world energy and drive rising GHG emissions, while renewable deployment has not kept pace with growing energy demand [1, 2].Solar and wind are presented as an economically viable route toward Paris Agreement climate goals.
Introduction: Materials discovery can take 15-20 years, spanning candidate selection, high-yield synthesis, and optimization for robust, reproducible devices [5,6].This timeframe leaves substantial room for improvement in sustainable-energy research.
Introduction: ML can predict material properties, generate structures with desired properties, identify renewable-energy patterns, and optimize energy management from devices to grids [8–10].These capabilities can reduce reliance on costly characterization and inform energy policy.
Introduction: The review introduces ML concepts and Acc(X)eleration Performance Indicators to assess accelerated energy-materials platforms.It also covers closed-loop ML frameworks, applications in energy harvesting, storage, and conversion, smart-grid integration, and future research directions.
Machine learning glossary: essential concepts: Large datasets and increased computing power have enabled diverse ML algorithms for energy problems, alongside a glossary of essential terminology [11,12].The review directs readers to prior reviews for detailed technical foundations [13–16].
Property prediction: Supervised learning predicts continuous or discrete properties from labelled data and can aid or replace physical simulations or measurements in some circumstances [17–19].A typical input-output formulation maps datapoint x to property y, such as band gap.
Generative materials design: Unsupervised generative models produce new examples from unlabelled data, while conditioning on physical properties enables inverse design toward improved properties [20–22].This property-to-structure approach biases generation toward desired outcomes.

Self-driving labs … Optimizing system management

ML supports a broad energy-research workflow, from closed-loop autonomous experimentation and characterization to computational acceleration and device- or grid-level energy-system management.

Self-driving labs: ML models enable autonomous laboratories to plan and perform experiments through automated retrosynthesis analysis, product prediction, and reaction-condition optimization [23].
Self-driving labs: Self-driving labs use ML to plan and perform experiments, including retrosynthesis, reaction-product prediction, and reaction-condition optimization, enabling closed-loop inverse design [21].Examples include reinforcement-learning-aided synthesis planning [24,25], convolutional neural networks for reaction prediction [26,27], and active-learning robotic workflows [23,28–33].
Empowering characterization: ML analyzes experimental observations by determining crystal structures from TEM images and identifying coordination environments, structural transitions, and crystal symmetry from spectroscopic and diffraction data.
Optimizing system management: ML supports energy-system management by predicting battery lifetimes, adapting to new building loads, and optimizing smart-grid control.Examples include battery-life prediction [40,41], long-short-term-memory building-load prediction [42], and reinforcement-learning smart-grid control [43].
Optimizing system management: ML-assisted management operates at both device and grid power levels, linking lifetime prediction, load adaptation, and performance optimization.

Acc(X)eleration Performance Indicators (XPIs) · Acceleration factor (AF) of new materials · Number of new materials with threshold performance

The paper proposes Acc(X)eleration Performance Indicators (XPIs) to provide comparable metrics for ML- and high-throughput-experimentation-accelerated materials discovery. The indicators quantify platform acceleration and the number of newly discovered materials exceeding a justified performance baseline.

Acc(X)eleration Performance Indicators (XPIs): Acc(X)eleration Performance Indicators (XPIs) address the lack of consistent comparators for accelerated materials discovery, complementing existing device-, plant-, and grid-level energy-management indicators.The authors propose XPIs because accelerated materials-discovery reports need a common baseline for evaluation and comparison.
Acc(X)eleration Performance Indicators (XPIs): ML for energy technologies shares methodologies and principles with biomedicine, but field-specific practice motivates metrics for evaluating and improving high-throughput-experimentation and ML workflows.The paper frames HTE and ML in materials discovery as a potential paradigm shift requiring systematic comparison.
Acc(X)eleration Performance Indicators (XPIs): Accelerated materials-discovery methods should ultimately be judged by commercialization time, although that metric is impractical for evaluating new platforms or rapidly choosing among them.Commercialization can require up to two decades, motivating substantially faster discovery workflows.
Acceleration factor (AF) of new materials: Acceleration factor (AF) is the ratio of materials synthesized and characterized per unit time using an accelerated platform versus traditional methods.An AF of 10 means the accelerated platform evaluates 10x more materials in the same period; for multiple target properties, report the rate-limiting AF.
Acceleration factor (AF) of new materials: The AF example shows that a value of 10 corresponds to evaluating 10x more materials than traditional methods over a given time period.For materials with multiple target properties, the reported AF should be limited by the slowest target-property evaluation.
Number of new materials with threshold performance: Number of new materials with threshold performance counts discoveries from an accelerated platform whose performance exceeds a defined baseline.The baseline must fairly represent the standard against which new materials are compared.
Number of new materials with threshold performance: For new perovskite solar-cell materials, the threshold metric can count devices whose performance surpasses the current record.The example illustrates why baseline selection is critical when assessing accelerated discovery outcomes.

Performance of best material over time … Closed-loop machine learning frameworks for materials discovery

The section defines performance indicators for evaluating accelerated materials-discovery platforms and contrasts traditional, computation-driven, ML-assisted, and automated virtual-screening workflows. It emphasizes performance growth, repeatability, resource requirements, and progressively more autonomous exploration of chemical space.

Performance of best material over time: Accelerated workflows should increase the best material’s absolute performance more rapidly over time than traditional methods, measured using application-specific metrics such as Faradaic or power-conversion efficiency.This XPI tracks the performance trajectory of the best material over time.
Repeatability and reproducibility of new materials: New materials should vary by no more than x% from their mean performance to ensure consistency and repeatability before inclusion in subsequent performance XPIs.This criterion helps screen out materials that might otherwise fail during commercialization.
Human cost of the accelerated platform: The accelerated platform’s human cost includes researcher hours for component design and ordering, infrastructure development, database maintenance, and platform operation.Reporting this total estimates the resources required to adapt an accelerated platform to other research programs.
Human cost of the accelerated platform: The XPIs can be measured across computational, experimental, and integrated systems, enabling consistent comparison of platform development.For sustainability, the robotic photocatalysis platform achieved an overall XPI score of 105.
Closed-loop machine learning frameworks for materials discovery: Traditional materials discovery uses an Edisonian trial-and-error loop: candidates are selected, synthesized and tested, and measured properties guide subsequent empirical searches.The workflow begins with a target application and candidate pool before iterative synthesis and characterization.
Closed-loop machine learning frameworks for materials discovery: ML-based virtual screening samples larger chemical-space regions by learning structure–property relationships and repeating the search when the desired material is not discovered.This approach is data-driven and aims to increase chemical-space coverage without adding equivalent time and effort.
Closed-loop machine learning frameworks for materials discovery: Automated virtual screening reduces reliance on human intuition by using computational and experimental data to train a generative ML model that directs the search after random initialization.The generative model addresses the inverse problem within an iterative computational–experimental discovery framework.

Photovoltaics materials discovery

Machine learning is accelerating photovoltaic materials discovery, particularly across perovskites’ vast chemical space, through representations that enable property prediction and self-driving experimentation. Effective screening must extend beyond bandgap to include defect density and stability, while laboratory automation can substantially reduce optimization experiments.

Atomic-feature representations enabled accurate perovskite-property prediction but omitted spatial relationships; image and graph representations address this limitation, with graphs accommodating varying system sizes.52,53,55,56,57,17Graph representations are especially suitable for organic-inorganic perovskites, whose crystal structures contain varying numbers of atoms.
Machine learning was used to predict lead-free perovskites with bandgaps appropriate for solar cells.54Bandgap prediction provides an important first screening step for photovoltaic materials.
Bandgap alone is insufficient for identifying useful optoelectronic materials because electronic defect density and stability are also important, while defect-energy datasets are computationally expensive to generate.Computational methods can address defect energies, but the cost of calculating structural defects limits dataset generation for ML training.
60 experiments instead of 500 optimized an organic solar cell using a self-driving laboratory, while robotic synthesis accelerated learning and reduced chemical costs.The approach demonstrates that ML can accelerate optoelectronic-material discovery even without a large experimental dataset.

Solar device structure and fabrication · Electrode and electrolyte materials design

Machine learning improves photovoltaic device design and manufacturing beyond the active layer, while electrochemical storage spans application-specific technologies including lithium-ion and redox flow batteries.

Solar device structure and fabrication: Photovoltaic performance requires optimizing layers beyond the active layer, including a top transparent conductive layer. [60,61]This layer must provide both high optical transparency and high electronic conductivity. [60,61]
Solar device structure and fabrication: A genetic algorithm optimized light-trapping topology, achieving 48.1% broadband absorption—more than threefold the Yablonovitch limit. [62]The transparent conductive layer must combine high optical transparency with high electronic conductivity. [60,61]
Solar device structure and fabrication: ML reduced yearly solar irradiance datasets to characteristic spectra, enabling optimal band-gap calculations under real-world conditions rather than a single standard spectrum. [63]Actual irradiance varies with solar position, atmospheric phenomena, and season. [63]
Solar device structure and fabrication: A CNN predicted current–voltage characteristics of as-cut silicon wafers from photoluminescence images. [64]
Solar device structure and fabrication: An artificial neural network predicted metallic front-contact resistance, a parameter critical to silicon solar-cell manufacturing. [65]
Electrode and electrolyte materials design: Electrochemical energy storage supports electric vehicles, consumer electronics, and stationary power stations through technologies with application-dependent efficacy.
Electrode and electrolyte materials design: Lithium-ion batteries provide excellent energy density for electronics and electric vehicles, whereas redox flow batteries have attracted attention for stationary power storage.

Developing new Li-ion battery materials that can deliver higher operating voltages, energy · Battery device and stack management

Machine learning is being applied both to discover higher-voltage battery materials and to manage battery lifetime, charging, storage profiles, and power under uncertainty. These approaches accelerate optimization, but generalization across chemistries and physically interpretable degradation models remain challenging.

Developing new Li-ion battery materials that can deliver higher operating voltages, energy: ML models use Materials Project data to predict electrode voltages for Na- and K-ion batteries and generate voltage-profile diagrams, supporting discovery of new battery materials [73].Layered oxides are widely used cathodes, but universal design principles for new alkali metal-ion battery materials remain undefined.
Developing new Li-ion battery materials that can deliver higher operating voltages, energy: ML explores electrolyte chemical space for organic redox-flow batteries, whose performance depends on active-material solubility and charge/discharge stability [76,77].S. Kim et al. proposed a multi-kernel-Ridge regression approach for this search.
Battery device and stack management: A mechanism-agnostic ML model predicts lithium-ion battery cycle life accurately at an early stage, overcoming limits of mechanism-specific capacity and power-loss models [40].Existing mechanistic and semiempirical models apply only to specific failure mechanisms or situations and cannot predict lifetime early.
Battery device and stack management: A combined early-life prediction and Bayesian-optimization model rapidly identifies charging protocols that maximize cycle life [41].ML can accelerate battery-lifetime optimization, although generalization to different chemistries remains unresolved.
Battery device and stack management: Hybrid physics-based ML models can improve degradation-model explainability and reduce overfitting by incorporating domain knowledge, but encoding battery degradation physics remains difficult [80].Although ML predicts battery lifetime, linking predictions to underlying degradation mechanisms and state of health remains challenging.
Battery device and stack management: Neural networks predict charge/discharge profiles for lithium iron phosphate and vanadium redox-flow stationary batteries, while reinforcement learning supports power management under environmental and application variability [82].Battery storage management must account for uncertainty and variability in both operating conditions and use cases.

Electrocatalyst materials discovery · Fuel cell and electrolyser device management

Machine learning accelerates electrocatalyst discovery by replacing costly calculations and simplifying reaction analysis, while supporting fuel-cell, electrolyser, and smart-grid performance management. These applications are constrained by complex, incompletely characterized systems but demonstrate substantial gains in control cost and performance.

Electrocatalyst materials discovery: Electrocatalyst discovery is difficult because activity depends on adsorption energies across many surface binding sites, with the number of possibilities increasing dramatically for alloys.Literature mining is further limited because publications may omit important variables, while synthesis and testing conditions can strongly affect performance.
Electrocatalyst materials discovery: Machine learning surrogate models reduce the cost of DFT-based electrocatalyst searches and can identify important reaction steps and likely pathways in complex mechanisms.The relevant mechanisms may involve hundreds of possible species and intermediates.
Electrocatalyst materials discovery: A closed-loop ML–DFT framework screened intermetallic chemical space for CO2 reduction and H2 evolution by predicting adsorption energies and verifying promising candidates with automatically selected DFT calculations.The workflow iteratively feeds verification results back into subsequent screening.
Electrocatalyst materials discovery: X-ray absorption spectroscopy combined with machine learning can help interpret measurements of electrocatalyst active sites and their evolution over time.The motivation is that XAS analysis relies heavily on human experience and expertise.
Fuel cell and electrolyser device management: Machine learning supports fuel-cell and electrolyser optimization, degradation and lifetime prediction, and fault detection across proton-exchange and solid-oxide devices.Reported applications include PEMFC degradation prediction, fault isolation using electrochemical impedance, solid-oxide leakage diagnosis, and electrolyser optimization for CO2/H2O reduction and chloralkali processes.
Incorporating XPIs: Acceleration factor is the most commonly reported energy-workflow performance indicator, but reporting additional indicators would clarify the time and human resources required to develop each platform.Authors generally report acceleration factor after completing platform development.
Integration of ML into smart power grids: Up to 80% lower operational cost was achieved when a relaxed deep-learning controller managed automatic generation control versus traditional heuristic strategies.A multi-agent reinforcement-learning strategy also improved control performance by ~10% versus other ML algorithms.
Integration of ML into smart power grids: Reinforcement-learning demand-side management reduced costs for both service providers and customers while dynamically shaping electricity consumption to balance renewable generation and load.Demand-side mechanisms include peak shaving, load growth, and load shifting.

Representing materials with novel geometries

Effective material representations capture inherent system properties and support downstream tasks such as transfer learning, visualization, attribution, and generative modeling. Designing general representations remains challenging because materials span molecular, periodic crystalline, and structurally complex systems.

Effective representations capture inherent system properties, including physical symmetries, while supporting transfer learning, visualization, attribution, and generative modeling.
Molecular materials have been represented with fingerprints, SMILES, SELFIES, and graphs, whereas crystalline materials additionally require periodicity.
Energy materials pose added representation challenges from large atom counts, specific symmetries, disorder or amorphous phases, defects and dislocations, and low dimensionality.
Self-supervised, multi-task, and meta-learning offer routes to better representations by leveraging synthetic tasks, correlations among properties, and adaptation to new or out-of-distribution datasets.

Data quality and more robust predictive models · Synthesizing and evaluating new materials for stability

Robust ML materials pipelines depend on larger, higher-quality, standardized datasets and models designed for low-data, uncertain, and out-of-distribution settings. Evaluating candidate materials also requires accounting for synthesis complexity and the long, multifaceted processes governing stability and degradation.

Data quality and more robust predictive models: Training-data size and quality are foundational because they determine predictive performance and the accuracy of discovered materials; deep learning scales more favorably with dataset size than traditional ML.
Data quality and more robust predictive models: Transparent, standardized reporting could support database consolidation, with NLP extraction [101], structured resources such as MatD3 [141], and autonomous laboratories [23,25] enabling continual updates.
Data quality and more robust predictive models: Low-data regimes require data-efficient models, active sampling [142], and data augmentation [143], while uncertainty quantification, interpretability, and regularization improve robustness.
Data quality and more robust predictive models: Generalizable models should maintain predictive performance for new material classes outside the original dataset’s distribution.
Synthesizing and evaluating new materials for stability: Formation energy estimates stability and synthesizability, but slightly positive values below a limit can indicate metastable phases whose synthesizability remains unclear [26,27].
Synthesizing and evaluating new materials for stability: Instead of directly predicting synthesis probability, ML can evaluate synthetic complexity using route accessibility or precedent reaction knowledge [156,157].Automated synthesis-planning algorithms are being developed for inorganic materials [158,159].
Synthesizing and evaluating new materials for stability: Material synthesis alone does not ensure commercialization because stability evaluation is necessarily long, and degradation can involve active-matter loss, inactive-phase growth, or defect propagation.Examples include rocksalt formation in layered Li-ion battery electrodes [160], Pt agglomeration in fuel cells [161], and cracking during battery cycling [162].

Optimizing the management of smart power grids · Bridging theory from materials to device to policy

ML can improve smart-grid decision-making and sustainable-energy development, but deployment is constrained by data scarcity, compliance and security concerns, and continued reliance on human judgment. The paper therefore advocates digital-twin-enabled workflows and a more integrated perspective spanning materials, devices, systems, and energy policy.

Optimizing the management of smart power grids: ML could automate smart-grid decisions for distributing dynamically supplied power more efficiently, but practical deployment remains difficult because of data scarcity and compliance and security concerns.166Capturing renewable-resource variation requires long-term data collection across peak/off-peak and seasonal conditions.167
Optimizing the management of smart power grids: Long-term renewable-resource data collection is used to represent peak/off-peak and seasonal variation, while risk-averse energy managers still rely on human decision-making despite ML’s intended treatment of uncertainty.167,168The passage states that these collections can span 24 hours to several years.
Optimizing the management of smart power grids: Digital twins can simulate physical-system dynamics and generate large amounts of high-quality synthetic data at relatively low cost for ML training [169,170].The digital twin may combine physical laws with ML models trained on samples from the physical system.
Bridging theory from materials to device to policy: Energy research should adopt a more integrated approach rather than focusing on one narrow aspect of larger problems.171Energy policy encompasses conversion, distribution, and utilization, while ML has been applied in energy economics and finance.
Bridging theory from materials to device to policy: ML applications in energy economics and finance include performance diagnostics and forecasting energy generation and consumption.The supplied examples include oil-well diagnostics and wind-power generation forecasting.
Conclusions: ML has contributed to advances across sustainable-energy research, from materials design through device management to system deployment.The paper identifies ML as particularly well-suited to discovering new materials and reports evidence that it can learn trends identified by human researchers over decades.

Figures

The figures depict ML-accelerated materials discovery, closed-loop design and characterization, optimized device management and fabrication, and opportunities across renewable-energy research. They also summarize representative advances, including candidate screening, improved lithium-ion conductor discovery, and electrocatalyst identification.

Research framework: The figures organize ML-enabled energy research around acceleration indicators, closed-loop materials design and characterization, and broader opportunities spanning renewable-energy harvesting, storage, conversion, and management.They include performance indicators, grand challenges, and opportunities for extensive ML-aided energy research.
Materials discovery: ML-guided screening reduced 12,831 candidate materials to 21 promising electrolytes, while another approach screened over 12,000 inorganic solids and identified four capable of suppressing Li dendrite growth.These examples illustrate the transition from conventional trial-and-error and high-throughput workflows toward ML-driven materials discovery.
Energy storage: The ML-guided search was 2.7 times more likely to identify fast Li-ion conductors and achieved at least a 44 times improvement in room-temperature Li-ion conductivity.Ab-initio molecular dynamics validated the clustering and identified top candidates, including 16 new Li-ion conductors.
Device management and fabrication: ML-based device-management policies explored power–performance trade-offs, reduced power relative to expert-based management, outperformed leading algorithms, and adapted as new data accumulated.The figures frame these advances within ML-optimized energy-device management and fabrication.