PaperlayerView exact revision on arXiv

Source-linked AI summary

Data-driven materials science: status, challenges and perspectives

Lauri Himanen, Amber Geurts, Adam S. Foster, Patrick Rinke

arXiv:1907.05644v2physics.comp-phcond-mat.mtrl-sci

TL;DR

Conventional materials discovery remains slow and costly, and the field faces challenges in data storage, standardization, and trustworthiness. This perspective reviews data-driven materials science and its infrastructures, concluding that the field combines mature high-throughput workflows with unresolved data-quality and integration challenges.

  • Problem

    Data-driven materials science still faces challenges in long-term storage and standardized, trustworthy materials data.

    Source: B. Overview of current materials… · p. 6, C. Current Challenges · p. 9, §VI · p. 12

  • Method

    The article reviews the historical development and current state of materials data infrastructures, alongside challenges in data organization, acquisition, quality, and machine learning.

    Source: §I · p. 2

  • Results

    High-throughput workflows have matured and are increasingly used to screen complex properties, while veracity remains the most problematic data-quality challenge.

    Source: §V · p. 12, §VI · p. 12

  • Takeaways & Limitations

    Materials data infrastructures and machine-learning approaches are established parts of materials research, but data organization and trustworthiness remain central challenges.

    Source: §I · p. 2, §VI · p. 12

Abstract

from arXiv · show

Data-driven science is heralded as a new paradigm in materials science. In this field, data is the new resource, and knowledge is extracted from materials data sets that are too big or complex for traditional human reasoning - typically with the intent to discover new or improved materials or materials phenomena. Multiple factors, including the open science movement, national funding, and progress in information technology, have fueled its development. Such related tools as materials databases, machine learning, and high-throughput methods are now established as parts of the materials research toolset. However, there are a variety of challenges that impede progress in data-driven materials science: data veracity, integration of experimental and computational data, data longevity, standardization, and the gap between industrial interests and academic efforts. In this perspective article, we discuss the historical development and current state of data-driven materials science, building from the early evolution of open science to the rapid expansion of materials data infrastructures. We also review key successes and challenges so far, providing a perspective on the future development of the field.

I. INTRODUCTION

Data-driven materials science applies large-scale data processing and machine learning to materials discovery, against a traditionally slow and costly discovery-to-deployment pipeline. This perspective traces the field’s open-science roots and infrastructure development, then reviews its challenges and future longevity.

I. INTRODUCTION

  • Materials discovery through conventional experimentation, theory, or computation remains slow, costly, and inefficient, with patent protection potentially ending before a material reaches market.The article describes this as a loss of proprietary advantage.Source: §I · p. 1
  • Data-driven materials science gathers available data in infrastructures and uses machine learning to discover materials, complementing the traditional discovery paradigms.The article frames it as the fourth scientific paradigm, following experimental, theoretical, and computational approaches.Source: Figure 1 · p. 1, §I · p. 2
  • The field combines automated data analysis with physical infrastructure that hosts and preserves data, requiring dedicated community effort and sustained investment.The article likens analysis to a search-and-find function and infrastructure to data storage.Source: §I · p. 2
  • Academia, industry, governments, and the public have different roles and interests in the data-driven materials science ecosystem.The article says this ecosystem’s vitality is crucial to the field’s success and longevity.Source: §I · p. 2
  • The article situates data-driven materials science within the open-science movement and traces infrastructure growth from databases to data centres and discovery platforms.It reviews current challenges, data organization, acquisition, quality, machine learning, and industrial perspectives on infrastructure longevity.Source: §I · p. 2

II. OPEN SCIENCE MOVEMENT

The Open Science movement combines accessible research, reproducibility, and digital collaboration. Open access to publications and data, alongside policy and funding requirements, has made openness a foundation for data-driven research.

II. OPEN SCIENCE MOVEMENT

  • Accessibility, public accountability, and reproducibility became central scientific norms, while the internet broadened access to research from preliminary data through final publications.The passage also notes continuing tensions over reward allocation and intellectual property.Source: §II · p. 2

A. The rise of Open Access publishing

  • Over 50% of newly published articles are Open Access, which is commonly provided through gold publication or green repository-deposit routes.The passage describes funder policies as one driver and estimates complete Open Access by 2040.Source: A. The rise of Open Access publishing · p. 3

B. Open Access data

  • Open Data developed from long-term data-centre efforts into a policy objective, while repositories proliferated and journals and funders increasingly required data sharing and management plans.The FAIR principles promote data that is searchable, accessible, flexible, and reusable.Source: B. Open Access data · p. 3

C. Open-source software for science

Open-source software supports collaborative scientific work, while materials data infrastructures evolved toward platforms for searching, mining, and discovering materials. The field’s development drew on early computational approaches and later data-science initiatives.

C. Open-source software for science

  • Open-source tools became a standard part of scientific work, with Git and GitHub supporting collaborative code updates, feedback, bug reports, and feature requests.The passage describes open-source software as available across many stages of scientific projects.Source: C. Open-source software for science · p. 3

III. MATERIALS DATA INFRASTRUCTURES

  • Materials data infrastructures collect, host, and provide materials data to stakeholders.The article turns from Open Science to early digital infrastructures and their current state.Source: §III · pp. 3–4

A. Development of materials infrastructures

  • Early materials informatics drew on CAL-PHAD, materials-property prediction, evolutionary algorithms, and high-throughput screening to support alloy design and narrow experimental searches.The Materials Project later used high-throughput computing to catalog known inorganic materials for interactive exploration and data mining.Source: A. Development of materials… · p. 4
  • Materials Genome Initiative-era infrastructures progressed from databases to data centres and then discovery platforms with workflows for searching, mining, and querying data.The article uses these as loose categories, with “materials data infrastructure” as the encompassing term.Source: A. Development of materials… · p. 4
  • Data-analysis capabilities created a feedback loop between hypothesis and experiment, including machine-learning tools that suggest candidate materials for further refinement.The article describes this as combining human trial-and-error with “artificial intuition.”Source: A. Development of materials… · p. 4

B. Overview of current materials infrastructures

Materials data platforms provide ways to distribute, store, and analyze data, while commercial services offer curated data and on-demand software environments. The section also notes differences in storage needs and data access options.

B. Overview of current materials infrastructures

  • Materials infrastructures and discovery platforms had become mature services, with an expanding range of features and services.The article notes that some infrastructures emerging at submission were not included in its platform feature tables.Source: B. Overview of current materials… · p. 5
  • Platforms provide browser-based exploration for accessible data use and APIs or downloadable databases for automated access to larger datasets.Full database downloads offer flexibility and performance but are impractical for large data volumes and can require SQL or ORM skills.Source: B. Overview of current materials… · pp. 5–6
  • Two recommended computational-data repositories guarantee storage for at least 10 years, while experimental data—especially imaging—poses greater challenges for central, free storage.Electron microscopes can generate tens of gigabytes in a day; experimental data storage is often hosted by universities or laboratories.Source: B. Overview of current materials… · p. 6
  • Commercial materials-informatics services sell access to curated proprietary data or provide on-demand cloud environments that avoid large hardware investments and specialized configuration.The passage identifies these services as potentially valuable to companies and research laboratories.Source: B. Overview of current materials… · p. 6

C. Current Challenges

Materials infrastructures face linked challenges in relevance, completeness, standardization, and stakeholder acceptance. Industry participation and long-term infrastructure support remain particular concerns.

C. Current Challenges

  • Adoption depends on relevance to different communities, including data volume, type, quality, completeness, and homogeneity, as well as useful analysis tools.Machine-learning tools are data hungry, and their integration is complicated by diverse approaches and data.Source: C. Current Challenges · p. 9
  • Most materials infrastructures contain mainly computational and almost no experimental data, limiting seamless comparison needed to validate predictions and support materials development.Computational data is more homogeneous and easier to curate, while promising materials still require evaluation, selection, and experimentation.Source: C. Current Challenges · p. 9
  • Standardized metadata is still an outstanding challenge; an ontology could provide a shared classification system and common language for materials data.Metadata relates data items, while an ontology can form the foundation for that metadata.Source: C. Current Challenges · p. 9
  • Acceptance also depends on user-friendly data upload and download, and trust supported by curation throughout the data lifecycle.Curation includes quality control from data creation and storage through archiving or deletion.Source: C. Current Challenges · p. 9
  • Industry engagement remains incomplete because companies need data different from what current platforms store, and academic–corporate R&D is often disconnected.Industry participation has not been systematically studied and is described as relying mostly on anecdotal evidence.Source: C. Current Challenges · p. 9
  • Funding for open-science platforms is increasing, but sustained operating support and long-term maintenance are not guaranteed.The article warns that infrastructures may become digital ruins when maintenance and usability remain secondary priorities.Source: C. Current Challenges · p. 9

IV. MATERIALS ONTOLOGY

A materials ontology can standardize classification and metadata, making data more searchable and shareable. The article reviews the practical burden and immaturity of ontology efforts while recommending unification wherever possible.

IV. MATERIALS ONTOLOGY

  • Raw materials data without metadata is neither searchable nor suitable for machine learning, so infrastructure design must select relevant materials and represent them completely.Metadata should also record sample synthesis and processing history or, for virtual samples, the generating code and computational settings.Source: §IV · p. 10
  • A materials ontology classifies materials, properties, units, limits, and their interrelations, defining search terms and supporting semantic reasoning for complex searches.It is also a shared standard for people working with the same data and a practical necessity in database design.Source: §IV · p. 10
  • Creating ontology machinery requires structures, encoding formats, standards, and widespread adoption; a standardized, complete materials ontology has not yet emerged.Competing efforts are less mature than those in fields such as biosciences, and companies may need internal domain-specific ontologies.Source: §IV · p. 10
  • File-based exchanges require interfacing software and expertise, while incompatible standards can cause conversion errors and data loss.Ontologies could standardize knowledge, though fitting data into them remains tedious and error-prone for humans.Source: §IV · p. 10
  • Figure 5 presents an example ontological hierarchy for structural characterization of materials.The figure is described as a materials tree of life.Source: Figure 5 · p. 11
  • OPTiMaDe is building a common interface across materials platforms, while the diversity of subfields may make one universal ontology impossible.The article recommends unifying ontologies where possible and disseminating those efforts to the community.Source: §IV · p. 11
  • Standardization can facilitate data sharing, and ontologies can reveal when materials or properties are missing from the represented scope.The authors note that incomplete coverage can lead a materials search tool to return irrelevant information.Source: §IV · p. 11

V. DATA CREATION

High-throughput experimental and computational methods generate the large, relevant datasets materials infrastructures need. The article describes advances in materials libraries, computational screening, standardized calculations, and automated workflows.

V. DATA CREATION

  • Experimental deposition and characterization advances, including combinatorial methods, enable rapid generation of composition–structure–property relationships.Composition and temperature gradients can be created across samples, supporting finer sampling and materials libraries that represent defined regions of materials space.Source: §V · p. 11
  • More than 560,000 computationally predicted materials have expanded the OQMD from roughly 30,000 experimentally known crystal structures.High-throughput workflows are also increasingly used to screen complex properties such as coupling and reorganization energies in organic crystals.Source: §V · pp. 11–12
  • Standardized computational specifications for DFT calculations make data compatible within and across databases.The Materials Project and AFLOW consortium specify details including k-point grid densities, cutoff energies, pseudopotentials, and convergence criteria.Source: §V · p. 12
  • Workflow managers connect computational steps and checks into graphs that create consistent datasets and support automated recovery from errors.Figure 6 illustrates a FireWorks workflow for creating an elastic-tensor dataset.Source: §V · p. 12, Figure 6 · p. 12
  • High-throughput experimental and computational methods provide the volume and automation needed to make materials infrastructures relevant and complete.The article concludes that data-creation bandwidth should not be an issue for future materials infrastructures.Source: §V · p. 12

VI. DATA QUALITY

Data quality matters for standardization and infrastructure acceptance, but veracity is especially difficult to quantify. Computational and experimental errors pose distinct challenges, while systematic data collection may support practical error estimation.

VI. DATA QUALITY

  • Veracity is the most problematic of the four Vs because trustworthiness is difficult to quantify, although data quality also affects standardization and infrastructure acceptance.The other Vs are volume, velocity, and variety; the text notes that volume and velocity may be addressed with more storage or faster processing.Source: §VI · pp. 12–13
  • Data veracity comprises bias, the offset from ground truth, and variance, the spread across identical runs; here these arise from model approximations or experimental uncertainty.The paper distinguishes this usage from bias and variance terminology commonly used in machine learning.Source: §VI · p. 12
  • Computational-data veracity is dominated by bias, whose estimation requires experimental data or higher-level theory and varies by chemical elements and material classes.Approximations may produce much larger errors for dispersion-governed, magnetic, strongly correlated, and heavy-element materials.Source: §VI · p. 12
  • Experimental bias can reflect incomplete knowledge of sample content or history and sample-environment interactions, while imperfections and equipment add variance that is hard to separate from bias.Reliable experimental data are available for simple systems such as small molecules or elemental crystals, but quality is harder to determine for more complex systems and measurements.Source: §VI · pp. 12–13
  • Because computational errors and experimental uncertainties jointly complicate veracity estimates, exhaustive experiment-computation comparisons are impractical across materials space, but extrapolation schemes may help.Extending such schemes to experimental-computational error estimation would require systematic data collection from both sources.Source: §VI · p. 13

VII. DATA ANALYTICS

Data platforms add tools, including machine learning, to operate on stored data and increase its value and relevance. The section introduces machine-learning approaches, their applications, and challenges of model transparency and algorithm choice.

VII. DATA ANALYTICS

  • Tools and apps operating directly on infrastructure databases, including machine learning, add value to stored data and enhance its relevance.The section reviews machine-learning types and how they could add value to materials infrastructures.Source: §VII · p. 13

A. Introduction to machine learning

  • Machine learning uses statistical models and optimization to find training-data patterns for predictions or decisions, with performance assessed by generalization to unseen data.Computers can handle larger, higher-dimensional data than human learners, according to the passage.Source: A. Introduction to machine learning · p. 13
  • Many statistical methods in machine learning predate the current AI boom, which the passage links to hardware, data gathering and curation, and method and library development.The passage identifies machine learning as a key ingredient of materials informatics.Source: A. Introduction to machine learning · p. 13
  • Supervised learning trains on input-output pairs; unsupervised learning works without outputs; reinforcement learning selects actions toward a long-term goal through exploration and exploitation.The passage describes supervised learning as the most mature and powerful subfield, and reinforcement learning as useful for exploring compound and material spaces.Source: A. Introduction to machine learning · p. 13
  • Figure 7 distinguishes data flow with white arrows from post-analysis actions that can be identified to improve model performance with green arrows.The figure presents key steps in building a machine-learning model.Source: Figure 7 · p. 14
  • Machine-learning models can be opaque because knowledge is encoded in potentially millions of parameters, and improved interpretability would benefit scientific understanding and wider acceptance.The passage contrasts scientific interpretability needs with commercial applications, where performance is more important.Source: A. Introduction to machine learning · p. 14

B. Specifics of machine learning in materials science

Materials-science machine learning combines data acquisition, feature engineering, model building, and performance assessment. Its success depends on matching features and algorithms to the problem and evaluating generalization against carefully designed data splits.

Workflow

  • Machine-learning applications commonly involve interdependent, iterative steps of data acquisition, feature engineering, model building, and analysis.These steps are illustrated in Fig. 7.Source: B. Specifics of machine learning in… · p. 14

Feature engineering

  • Because raw data is usually unsuitable for machine learning, feature engineering extracts informative inputs and can encode known symmetries in atomic geometries.Feature learning can instead discover features from data, while analyzing model features can help verify decisions and interpretability.Source: B. Specifics of machine learning in… · pp. 14–15

Model selection

  • No single algorithm is optimal for all problems, so model choice depends on the task, data size, and input structure.Custom neural-network architectures for atomistic geometries can reduce the need for feature engineering.Source: B. Specifics of machine learning in… · p. 15
  • Kernel ridge regression becomes costly on large datasets because brute-force matrix inversion scales as O(n^3), while neural networks can train in small batches.Gaussian processes, Bayesian optimization, subgroup discovery, compressed sensing, and LASSO or SISSO are among tools suited to small datasets.Source: B. Specifics of machine learning in… · p. 15

Performance assessment

  • Performance assessment checks whether models generalize to unseen data, but clustered samples can make random train-test splits yield unrealistic metrics.Leave-one-cluster-out cross-validation can provide more realistic performance metrics when material types form clusters.Source: B. Specifics of machine learning in… · p. 15

Challenges

  • Limited data, few benchmark sets, nonstandard performance metrics, and insufficiently interpretable decisions remain challenges for materials-science machine learning.The field's rapid development also makes selecting data, features, models, and analysis tools difficult; careful curation and standardization can mitigate this to some degree.Source: B. Specifics of machine learning in… · p. 16

VIII. APPLICATIONS

The article reviews data-driven applications in materials discovery, understanding materials phenomena, and advancing materials modelling. It describes screened and experimentally synthesized materials, data-derived physical insights, and surrogate models that can reduce computational cost.

Materials discovery

  • Forward screening has helped identify novel or improved materials for experimental synthesis, with examples including OLED molecules, polymer dielectrics, Heusler structures, and alloys.Other screened candidates, including photovoltaics and topological insulators, still await experimental validation.Source: §VIII · p. 16
  • Inverse materials design remains difficult in practice because generating valid, synthesizable candidates with desired properties is challenging.Forward screening also cannot find a desired material absent from the candidate pool.Source: §VIII · p. 16

Understanding materials phenomena

  • Data-driven methods can help reveal materials phenomena by deducing physical laws and identifying interpretable descriptors of mechanisms.Compressed sensing has identified descriptors controlling perovskite and monolayer metal-oxide stability; unsupervised maps have supported studies of superconductors, molecules, phase transitions, and catalysis.Source: §VIII · pp. 16–17

Advancing materials modelling

  • Surrogate models trained on accurate data can reproduce complex chemical phenomena at lower computational cost while sacrificing some accuracy.Machine-learned force fields promise near-DFT accuracy at costs closer to classical force fields and have supported studies of amorphous-carbon growth and aqueous nanoclusters.Source: §VIII · p. 17
  • Learning coupled-cluster energies and training exchange-correlation functionals or potential-to-density mappings can extend theoretical knowledge to new time and size regimes.The cited efforts concern very accurate but computationally expensive coupled-cluster calculations and density- or wave-function-based methods.Source: §VIII · p. 17

IX. STAKEHOLDER RELATIONS

Data-driven materials science depends on cooperation among academic, industrial, governmental, and public stakeholders, yet the field's ecosystem and industry participation remain limited. The article describes funding gaps, industry-facing data models, and tensions around proprietary data, skills, and research-industry collaboration.

Ecosystem and stakeholders

  • The field remains scattered and lacks a supportive ecosystem with non-academic stakeholders, while wider stakeholder support is needed for infrastructure diffusion and longevity.The article frames academia, government, industry, and the public as actors involved in developing the field.Source: §IX · p. 17

Funding

  • Targeted funding for materials-science and industry applications has been limited outside notable national initiatives.The article cites US Materials Genome Initiative funding, EU projects facilitated through high-performance-computing grants, and support for Switzerland's MARVEL centre.Source: §IX · p. 17

Industry adoption

  • Industry has remained reluctant to invest in data-driven materials research, with possible contributing factors including proprietary-data interests, long materials-development timelines, and limited informatics expertise.Firms' R&D data may be undocumented and intangible, requiring structuring, connection, and updating before proprietary databases can be created.Source: §IX · p. 18

Industry-platform relationships

  • Firms can collaborate with existing data platforms, build proprietary digital infrastructures, or use intermediaries that provide curated materials data and related services.Intermediaries include materials-informatics companies that can help firms develop skills, capabilities, and knowledge.Source: §IX · p. 18

Ecosystem tensions

  • Commercial opportunities may encourage private data ownership that keeps valuable data from the scientific community, while industry recruitment raises concerns about academic brain drain.Institutionalizing data science and machine learning in education could increase research funding, students, and industry interest or collaboration.Source: §IX · p. 19

Research-industry collaboration

  • Promising materials identified through informatics still require evaluation, experimentation, certification, and manufacturing, making computational-experimental synergies important in firms.The field's broader establishment is framed as requiring joint efforts among research, industry, public, and governmental organizations.Source: §IX · p. 19

X. TAKE-HOME MESSAGES

The review surveys the current state of data-driven materials science and identifies persistent challenges despite the field's historical maturation.

Take-home messages

  • The field has matured greatly from a historical perspective, but relevance, completeness, standardization, acceptance, and longevity remain key challenges.The article presents an overview of the current state of data-driven materials science.Source: §X · p. 19
  • The review focuses on the current state of data-driven materials science.Its stated scope is an overview of the field.Source: §X · p. 19
  • Relevance, completeness, standardization, acceptance, and longevity are identified as challenges still needing attention.These concerns qualify the field's historical maturation.Source: §X · p. 19

Related papers

Loading 1907.05644v2…