1. Three-line summary, prerequisites, and where this fits in the series

• MI is a tool for narrowing down what to try next. It turns a material's composition, structure and process into numbers (descriptors), learns trends from past data with machine learning, and narrows where to look. It handles both the “forward problem” (predict properties from a material) and the “inverse problem” (find a material from the properties you want)
• Computation, experiment and AI share the work. Computation (DFT and the like) “produces candidates cheaply and fast,” experiment “checks whether it is really true,” and AI “narrows where to look.” The three form a loop connected by data
• The one point to remember: a prediction, or a computed stability, is a candidate, not a discovery. Only after a material has been made, measured and reproduced can its property be treated as a fact

2. Searching for materials is a treasure hunt on a very large map

(1) An analogy: looking for the treasure

Think of the search for a new material as looking for treasure on a very large map. The map is the whole set of candidate materials. The treasure is a material with the property you want. Digging one spot (one synthesis and measurement) costs time and money. The map is so large that you cannot dig everywhere.

How large is the map? Davies and colleagues estimated that, taking the first 103 elements of the periodic table, the number of four-component compounds, counting the element ratios as well, exceeds one trillion (10¹²). If you then remove chemically implausible compositions using valency and electronegativity (a rough measure of how strongly an atom pulls electrons), about ten billion (10¹⁰) remainSourced [Ref. 4].

To get a feel for ten billion, suppose we test them all at one per day and simply divide: 10¹⁰ ÷ 365.25 ≈ 27 million years (Our calculation). Even at one per hour without a break it takes about 1.1 million years, and even a high-throughput setup (making and measuring many samples in parallel) at 1,000 per day would take about 27,000 years. This is what the right-hand side of Fig. 2 shows.

Our calculation. The figure of 10¹⁰ is the estimate of Davies et al. [Ref. 4]; the division is ours. It ignores success rates, labor and the number of instruments, and is only meant to convey orders of magnitude.

(2) Time: 10 to 20 years before a new material is used

The other barrier is time. The white paper on the Materials Genome Initiative (MGI), issued by the U.S. National Science and Technology Council in June 2011, says it takes typically about 10 to 20 years for a new class of materials to move from initial research to first use. As an example it cites the lithium-ion battery, which took 20 years from a laboratory concept proposed in the mid-1970s to wide use in the late 1990sSourced [Ref. 1]. The current MGI website says it “can take 20 or more years” to move a material from discovery to marketSourced [Ref. 3].

The white paper names, as one reason for the slowness, continued reliance on scientific intuition and trial-and-error experiments. It also notes that the seven stages a material passes through on the way to market are handled by different teams, often at different organizations, with few chances for feedback between stages [Ref. 1]. Against this, MGI aims to discover, develop, manufacture and deploy advanced materials at least twice as fast, at a fraction of the costSourced [Ref. 1]. That is a target; whether it has been achieved is a separate question (Chapter 10).

3. What materials informatics (MI) is

Materials informatics (MI) combines materials science with information science, which handles data, to speed up materials research and development. The review by Ramprasad and colleagues describes its central method as follows: build “surrogate” machine-learning models that predict quickly, based purely on past data. A material is usually first expressed as a set of numbers (a fingerprint; in this series, a descriptor), and a mapping learned from data links it to the property of interestSourced [Ref. 5].

Here are the three terms that just appeared, in the wording we use throughout the series.

• Machine learning = an approximator that learns trends from past data. It fits the relation between inputs and outputs, in the form of a formula, to the data. Limit of the analogy: it is weak outside the range of the data (extrapolation), and it does not understand physical laws
• Descriptor = a “business card” for a material, written in numbers. It turns composition, structure and process into values a machine can read (the next article, on materials data and descriptors, covers it in detail)
• DFT (density functional theory) = a physics simulation that calculates a material's properties from the behavior of its electrons. It contains approximations and does not necessarily agree with measurements (see our explainers on quantum chemistry and materials simulation, and on the fundamentals of density functional theory)

Agrawal and Choudhary discuss the search for materials in large datasets as part of the “fourth paradigm” of science (data-driven discovery), after experiment, theory and computation, and present examples of MI as “forward models” that predict properties and “inverse models” that find candidate materialsSourced [Ref. 6]. MI can be seen as the field that applies this idea to materials (commentary).

4. Forward and inverse problems

The questions of MI come in two directions. The forward problem asks: “For this material (composition, structure, process), what will the properties be?” The inverse problem asks: “I want this property. What material should I make?”

View original SVG, values and sources

Original SVG reference: Conceptual diagram (vector drawing). Left: the directions of the forward and inverse problems. Right: a made-up case where the property rises and falls, so two conditions reach the target (dotted line). Real materials, with combinations of composition, structure and process, can have many more answers.

The inverse problem is harder because the answer is not unique. In the figure on the right, for a material variable whose property rises and then falls, two conditions satisfy the same target property. For real materials, combinations of composition, structure and process can give many more answers, or none at all (commentary).

Zunger discusses inverse design as a new style of collaboration between theory and experiment: declare the desired functionality first, use theoretical calculations to predict which stable and synthesizable compounds show it, and improve the predictions through repeated cycles of prediction, synthesis and characterizationSourced [Ref. 7]. Sanchez-Lengeling and Aspuru-Guzik review “inverse design” of molecules with desired properties using generative machine-learning models [Ref. 8] (generative models are covered in our explainer on generative AI for materials design).

5. How computation, experiment and AI share the work

MI does not stand on AI alone. Computation, experiment and AI each take on what they do best.

View original SVG, values and sources

Original SVG reference: Conceptual diagram (vector drawing). Roles of computation, experiment and AI. Arrows show where candidates and data travel; this is our summary, not the setup of a particular lab.

This table is our own commentary. Each row describes a general tendency and does not apply to every method.

Let us look at “computation does not necessarily match measurement” in numbers. The paper on the Open Quantum Materials Database (OQMD) compares DFT-calculated formation energies (the energy change when a compound forms from its elements) with measured values for 1,670 compounds. The mean absolute error is 0.096 eV/atom. It also finds that the mean absolute error between experiments (different reports measuring the same compound) is 0.082 eV/atom, a “surprisingly large” value, and argues that a significant fraction of the error between DFT and experiment may be attributed to experimental uncertaintySourced [Ref. 9].

6. The basic flow of MI

MI work divides roughly into six steps. Here is what each one does, briefly. Fig. 6 also shows which article in this series covers each step in more detail.

1. Define the goal. Write in one line which property you want to know, in what range, with what accuracy, and by when. If the goal is vague, nothing after it can be decided
2. Collect data. Gather it from your own laboratory notebooks and measurements, and from public databases (our explainer on materials databases and the FAIR principles)
3. Turn it into numbers. Convert each material into a descriptor (a business card) and line them up in a table, one row per sample (our explainer on materials data and descriptors)
4. Build a model. Build a predictor that learns trends from the data. Materials data are often scarce, so overfitting (memorizing random fluctuations) needs care (our explainer on machine learning with small data)
5. Predict, with uncertainty. Look not only at the predicted value but also at how far to trust it (our explainer on uncertainty and applicability domain)
6. Check by experiment. Make and measure the candidates you narrowed down. The results return as new data to step 2 (our explainers on Bayesian optimization and active learning, and on self-driving laboratories)

7. How much does it help? The effect of narrowing down

How much do you gain by “narrowing where to look”? A small numerical experiment answers this (Our calculation).

• Setup. There are 10,000 candidates, and the truly good ones (true property in the top 1%, that is 100) count as “hits.” You can only test 100
• Picking 100 at random. The expected number of hits is 100 × 1% = 1
• Narrowing by prediction. A predictor gives each candidate a score. Call the correlation between this score and the true value ρ (rho). ρ is a number between 0 and 1; the closer to 1, the more closely the ranking by score follows the ranking by true value. We pick the top 100 by score and count how many hits it contains (average of 500 trials)

The results are in Fig. 7. With a useless predictor, ρ = 0, there are about 1.1 hits, the same as picking at random. At ρ = 0.5 there are about 12.9, and at ρ = 0.9 about 54.2. If you can test only 20, the random expectation is 0.2 hits; ρ = 0.5 gives about 3.9 and ρ = 0.9 about 17.1.

Our calculation. Synthetic data (true property normally distributed; the prediction score is a random number with correlation ρ to the true value). These are not data on real materials. The calculation assumes that the quality of prediction (ρ) holds for new candidates.

8. A short history, and how to read “AI discovered new materials”

Here is a timeline of MI, limited to what we could confirm in papers and official documents (Fig. 8).

Three currents can be read from the timeline (commentary). (1) Policy. In 2011 the MGI put data sharing and computation at the center of materials development. The 2021 strategic plan sets three goals for the next five years: unify the Materials Innovation Infrastructure, harness the power of materials data, and educate and train the workforce [Ref. 2]. (2) Data infrastructure. In 2012–13, AFLOW [Ref. 16], the Materials Project [Ref. 17] and OQMD [Ref. 18] were published, making DFT-computed materials data widely usable (see our explainer on materials databases and the FAIR principles). (3) Evaluation. Since 2020, frameworks that compare methods under common conditions have emerged, such as Matbench [Ref. 22] and Matbench Discovery [Ref. 23] (Chapter 11).

(1) A case: how to read “2.2 million structures”

One study often reported as “AI discovered new materials” is GNoME (Graph Networks for Materials Exploration) by Merchant et al. The abstract says that graph neural networks (AI that takes the connections between atoms as input) trained at scale enabled the discovery of 2.2 million structures below the convex hull of previously known stable crystals (the energy boundary of stability). The same abstract says that, of the stable structures, 736 have already been independently experimentally realizedSourced [Ref. 12].

On the other hand, Cheetham and Seshadri examined the claims of that paper in a perspective in Chemistry of Materials. They write that they find “scant evidence for compounds that fulfill the trifecta of novelty, credibility, and utility.” They also say the methods appear to hold promise and that there is a great need to incorporate domain expertise in materials synthesis and crystallographySourced [Ref. 13].

9. What a materials engineer can do starting today

You can start MI without installing a large system; start by organizing your own data. These six things can be done by a materials engineer today (commentary).

This table is our own commentary. The Raccuglia et al. study in row 3 concerned hydrothermal synthesis of vanadium selenites (records from laboratory notebooks) and does not necessarily carry over to other materials.

10. Limits, what we could not confirm, and a summary

(1) What we could not confirm

• Origin of the term. We only checked the bibliographic details of Rajan's paper “Materials informatics” (Materials Today, 2005) and did not read the full text. We did not confirm whether it is the first use of the term
• Progress of MGI. The 2021 strategic plan says the MGI approach has already led to accelerated deployment of advanced materials “in select applications” [Ref. 2]. We did not confirm whether the whole target of twice the speed at a fraction of the cost has been met. Strategies of countries other than the U.S. are covered in our explainer on national MI strategies and data infrastructure
• Time figures. “10 to 20 years” and “20 or more years” are statements in MGI documents [Refs. 1, 3]. They vary widely by material and application, and we did not verify them independently
• GNoME. We do not assess the novelty or usefulness of the predicted structures. We set the claims of the paper [Ref. 12] and the perspective [Ref. 13] side by side
• Efforts in Japan. Domestic activity such as NIMS is outside the scope we checked here (other explainers cover it)
• Text of standards. For ISO/IEC standards we only checked number, title and publication date, not the full text
• An MI-specific international standard. We could not confirm one (Chapter 11)
• Numerical experiments. The calculations in Chapters 2 and 7 are a plain division and synthetic data. They do not show how well predictions would work for real materials

11. How to read the yardsticks and standards

How good MI is depends on which yardstick you use. The typical yardsticks are prediction error (MAE, mean absolute error; RMSE; R²), how the data were split (randomly, or by composition or time), benchmarks (rankings on common tasks), the difference between computation and measurement, and the success rate and reproducibility of experimental confirmation. Each one needs care about what it measures and what it does not. For example, a small prediction error does not mean the model is right outside the range of its training data. The table below lists the benchmarks and the standards and principles related to MI that we could confirm (checked 22 October 2026). We could not confirm an international standard specific to MI.

The arrangement of the table is Our calculation (organization). ISO/IEC numbers, titles and dates were checked on the official ISO pages (22 October 2026). Standard texts are paywalled, so we checked only titles and dates and wrote no requirements.

Cautions in reading (for materials engineers)

• A benchmark ranking differs from performance on your material. A benchmark is a ranking for specific data, splits and metrics. When you use it for your material, measure again on your data, with a split that matches your purpose
• Separate “prediction” from “experimental confirmation.” A small prediction error is not the same as a property confirmed by synthesis. For experimental confirmation, look at the number of samples, the definition of success, reproduction, and whether anyone verified it independently
• Data leakage and the split trap. If samples of the same series fall in both training and evaluation, the number is overstated (see our explainer on the basics of machine learning and neural networks)
• Most principles and standards are recommendations. The FAIR principles and the NIST AI RMF are guidance without legal force. When something says “compliant,” check which requirement of which document is meant
• A plan is not an achievement. MGI's target of twice the speed is a target. Evidence of achievement has to be checked separately

12. Glossary

13. References (primary sources)

1. National Science and Technology Council "Materials Genome Initiative for Global Competitiveness", June 2011 — mgi.gov/sites/mgi/files/materials_genome_initiative-final.pdf
2. NSTC Subcommittee on the Materials Genome Initiative "Materials Genome Initiative Strategic Plan", November 2021 — mgi.gov/sites/mgi/files/MGI-2021-Strategic-Plan.pdf
3. Materials Genome Initiative "About the Materials Genome Initiative" (official website) — mgi.gov/about-materials-genome-initiative
4. Davies D. W. et al. "Computational Screening of All Stoichiometric Inorganic Materials", Chem 1, 617–627 (2016), DOI 10.1016/j.chempr.2016.09.010 — doi.org/10.1016/j.chempr.2016.09.010
5. Ramprasad R. et al. "Machine learning in materials informatics: recent applications and prospects", npj Comput. Mater. 3, 54 (2017), DOI 10.1038/s41524-017-0056-5 — doi.org/10.1038/s41524-017-0056-5
6. Agrawal A., Choudhary A. "Perspective: Materials informatics and big data: Realization of the 'fourth paradigm' of science in materials science", APL Mater. 4, 053208 (2016), DOI 10.1063/1.4946894 — doi.org/10.1063/1.4946894
7. Zunger A. "Inverse design in search of materials with target functionalities", Nat. Rev. Chem. 2, 0121 (2018), DOI 10.1038/s41570-018-0121 — doi.org/10.1038/s41570-018-0121
8. Sanchez-Lengeling B., Aspuru-Guzik A. "Inverse molecular design using machine learning: Generative models for matter engineering", Science 361, 360–365 (2018), DOI 10.1126/science.aat2663 — doi.org/10.1126/science.aat2663
9. Kirklin S. et al. "The Open Quantum Materials Database (OQMD): assessing the accuracy of DFT formation energies", npj Comput. Mater. 1, 15010 (2015), DOI 10.1038/npjcompumats.2015.10 (open access) — doi.org/10.1038/npjcompumats.2015.10
10. Ling J. et al. "High-Dimensional Materials and Process Optimization Using Data-Driven Experimental Design with Well-Calibrated Uncertainty Estimates", Integr. Mater. Manuf. Innov. 6, 207–217 (2017), DOI 10.1007/s40192-017-0098-z — doi.org/10.1007/s40192-017-0098-z
11. Li K. et al. "Exploiting redundancy in large materials datasets for efficient machine learning with less data", Nat. Commun. 14, 7283 (2023), DOI 10.1038/s41467-023-42992-y (open access) — doi.org/10.1038/s41467-023-42992-y
12. Merchant A. et al. "Scaling deep learning for materials discovery", Nature 624, 80–85 (2023), DOI 10.1038/s41586-023-06735-9 — doi.org/10.1038/s41586-023-06735-9
13. Cheetham A. K., Seshadri R. "Artificial Intelligence Driving Materials Discovery? Perspective on the Article: Scaling Deep Learning for Materials Discovery", Chem. Mater. 36, 3490–3495 (2024), DOI 10.1021/acs.chemmater.4c00643 — doi.org/10.1021/acs.chemmater.4c00643
14. Rajan K. "Materials informatics", Materials Today 8(10), 38–45 (2005), DOI 10.1016/S1369-7021(05)71123-8 — doi.org/10.1016/S1369-7021(05)71123-8
15. Fischer C. C. et al. "Predicting crystal structure by merging data mining with quantum mechanics", Nature Materials 5, 641–646 (2006), DOI 10.1038/nmat1691 — doi.org/10.1038/nmat1691
16. Curtarolo S. et al. "AFLOW: An automatic framework for high-throughput materials discovery", Comput. Mater. Sci. 58, 218–226 (2012), DOI 10.1016/j.commatsci.2012.02.005 — doi.org/10.1016/j.commatsci.2012.02.005
17. Jain A. et al. "Commentary: The Materials Project: A materials genome approach to accelerating materials innovation", APL Mater. 1, 011002 (2013), DOI 10.1063/1.4812323 — doi.org/10.1063/1.4812323
18. Saal J. E. et al. "Materials Design and Discovery with High-Throughput Density Functional Theory: The Open Quantum Materials Database (OQMD)", JOM 65, 1501–1509 (2013), DOI 10.1007/s11837-013-0755-4 — doi.org/10.1007/s11837-013-0755-4
19. Wilkinson M. D. et al. "The FAIR Guiding Principles for scientific data management and stewardship", Sci. Data 3, 160018 (2016), DOI 10.1038/sdata.2016.18 (open access) — doi.org/10.1038/sdata.2016.18
20. Raccuglia P. et al. "Machine-learning-assisted materials discovery using failed experiments", Nature 533, 73–76 (2016), DOI 10.1038/nature17439 — doi.org/10.1038/nature17439
21. Butler K. T. et al. "Machine learning for molecular and materials science", Nature 559, 547–555 (2018), DOI 10.1038/s41586-018-0337-2 — doi.org/10.1038/s41586-018-0337-2
22. Dunn A. et al. "Benchmarking materials property prediction methods: the Matbench test set and Automatminer reference algorithm", npj Comput. Mater. 6, 138 (2020), DOI 10.1038/s41524-020-00406-3 (open access) — doi.org/10.1038/s41524-020-00406-3
23. Riebesell J. et al. "A framework to evaluate machine learning crystal stability predictions", Nat. Mach. Intell. 7, 836–847 (2025), DOI 10.1038/s42256-025-01055-1 — doi.org/10.1038/s42256-025-01055-1
24. Choudhary K. et al. "JARVIS-Leaderboard: a large scale benchmark of materials design methods", npj Comput. Mater. 10, 93 (2024), DOI 10.1038/s41524-024-01259-w — doi.org/10.1038/s41524-024-01259-w
25. VAMAS "Versailles Project on Advanced Materials and Standards" (official website) — vamas.org
26. ISO/IEC "ISO/IEC 22989:2022 Information technology — Artificial intelligence — Artificial intelligence concepts and terminology" (ISO page) — iso.org/standard/74296.html
27. ISO/IEC "ISO/IEC 23053:2022 Framework for Artificial Intelligence (AI) Systems Using Machine Learning (ML)" (ISO page) — iso.org/standard/74438.html
28. ISO/IEC "ISO/IEC 5259-1:2024 Artificial intelligence — Data quality for analytics and machine learning (ML) — Part 1: Overview, terminology, and examples" (ISO page) — iso.org/standard/81088.html
29. ISO/IEC "ISO/IEC 5259-3:2024 Artificial intelligence — Data quality for analytics and machine learning (ML) — Part 3: Data quality management requirements and guidelines" (ISO page) — iso.org/standard/81092.html
30. ISO/IEC "ISO/IEC 42001:2023 AI management systems" (ISO page) — iso.org/standard/81230.html
31. NIST "Artificial Intelligence Risk Management Framework (AI RMF 1.0)", NIST AI 100-1, January 2023, DOI 10.6028/NIST.AI.100-1 — doi.org/10.6028/NIST.AI.100-1

14. Claim-to-source audit

Last updated: 22 October 2026 / Sources are limited to primary materials (official U.S. MGI documents, peer-reviewed papers, and official pages of ISO, NIST and VAMAS); no market-research estimates are used. The full text of ISO/IEC standards is paywalled and was not checked. The narrowing-down numerical experiment is a calculation on synthetic data we generated, not a result on real materials. We could not confirm progress toward MGI's goal, the full text of Rajan's paper, a judgment on the novelty or usefulness of GNoME, or an MI-specific international standard, so these are not stated. All figures are explanatory concept diagrams or drawings that include our own calculation; the hero image and Fig. 9 are ImageGen-generated conceptual illustrations.