1. The short version / before you start / where this fits in the series

• Electronic materials are a field of active MI research: properties such as band gap and dielectric constant are easy to compute (with DFT and similar methods), and databases are well developed. The dielectric database has 1,056 compounds; when it was published it was larger than the number of compounds with a measured dielectric constant (a few hundred) Sourced [Reference 7].
• Write calculated and measured values in separate columns: in an internal test by the Materials Project, band gaps from commonly used calculations (GGA/PBE) were about 40% lower than measurements on average, and an MAE of 0.6 eV remained even after rescaling by a factor of 1.6 Sourced [Reference 5]. A machine-learning result that matches PBE values with an error of 0.16 eV does not mean the measured values were matched [Reference 6].
• Few computationally screened candidates have been confirmed by experiment: for 2D materials, the paper says that only a few dozen 2D materials have been synthesized or exfoliated, against 1,036 compounds calculated to be easily exfoliable Sourced [Reference 11]. A review of thermoelectric materials says that, although machine learning reports high prediction accuracy, only a handful of predictions led to the experimental discovery of new high-performance materials [Reference 17].

2. Properties of electronic materials and the role of calculation

The properties of electronic materials are set by the behaviour of electrons and atoms inside the crystal. So properties can be estimated by calculation before the material is made. This "can be estimated first by calculation" is probably one reason why MI research is active in electronic materials (commentary). Let us briefly explain the four properties treated here.

• Band gap: the energy difference needed before electrons can move, in eV (electronvolts). It separates semiconductors from insulators and sets the wavelengths of light absorbed or emitted.
• Dielectric constant: how much charge shifts (polarizes) inside a material when an electric field is applied. It matters when choosing materials for capacitors and transistor insulating films.
• Thermoelectric properties: the ability to produce a voltage from a temperature difference (or to move heat with a voltage). The performance index zT is set by the Seebeck coefficient (voltage per temperature difference), electrical conductivity and thermal conductivity, and is larger for materials that conduct electricity well and heat poorly.
• Ease of exfoliating 2D materials: the energy (binding energy) needed to peel an atomically thin sheet off a layered crystal.

As the last column of the figure shows, the relation between calculation and measurement differs from property to property. From the next chapter we check, property by property, the gap between calculated and measured values in primary sources.

3. Band gap: calculated, measured and machine-learned values

(1) Calculated values tend to come out smaller than measured ones

The band gap is one of the properties most often handled in MI. But commonly used DFT calculations (LDA, GGA/PBE and so on) tend to underestimate the band gap. The official Materials Project documentation says that gaps computed with these methods are "severely underestimated", that the disagreement is reported to be around 50% in the literature, that internal testing by the project finds an underestimate of about 40%, and that some known insulators are predicted to be metallic Sourced [Reference 5].

A figure in the same documentation shows that, in a test of 237 compounds, computed values were smaller than measured ones by an average factor of about 1.6, and that even after accounting for this shift an MAE of 0.6 eV remains [Reference 5].

An example to build a feel for the numbers: for a material whose measured gap is 1.6 eV, a computed value smaller by the average factor of 1.6 would be about 1.0 eV Our calculation (this is an average; individual materials can differ greatly, and the post-correction MAE of 0.6 eV is a guide to that scatter). A 1 eV gap and a 1.6 eV gap use completely different wavelengths of light.

The choice of functional (the approximation used) also changes the result. Borlido and co-workers assessed 12 functionals, and in a follow-up study 21 functionals, using experimental structures and band gaps of 472 non-magnetic materials (including covalent, ionic and van der Waals solids), and report that mBJ, HSE06 and HLE16 are the most accurate Sourced [References 2, 3]. So even "calculated by DFT" has error sizes that depend on the functional used. When reading a paper, check the name of the functional.

(2) Which "value" is machine learning matching?

Many studies predict the band gap by machine learning. What matters is what the training labels (the values learned as "correct answers") are. If the labels are DFT values, the model learns to match "DFT values". If the labels are measured values, it learns to match "measured values".

Zhuo and co-workers built a machine-learning model that predicts the band gap from composition alone using only experimentally measured band gaps for training. It first classifies metals and non-metals, and then regresses the gap of the non-metals. The authors report that the predicted values are significantly closer to experimental values than PBE-level DFT values Sourced [Reference 10].

Matbench has two tasks that express this difference: a task that predicts the "measured gap from composition alone" (data of Zhuo and co-workers, 4,604 records) and a task that predicts the "gap calculated with PBE from the crystal structure" (Materials Project data, 106,113 records) Sourced [References 1, 6].

On the official leaderboards (checked 22 October 2026), the MAE on the measured task was 1.1435 eV for Dummy (the baseline that answers the mean) and 0.2865 eV for the best method (Darwin). On the PBE task, Dummy was 1.3272 eV and the best method (coGN) was 0.1559 eV Sourced [Reference 6]. The error on the PBE task is small because the labels are tidy values produced by calculation. The labels of the measured task are a collection of values that were actually measured, and they contain scatter caused by different measurement conditions (up to this point, commentary). According to the Matbench description, in the measured task, compositions whose reported gaps disagree by more than 0.1 eV were removed [Reference 6]. That is, measured values themselves disagree with one another in practice.

4. Dielectrics: the values are off, but can the ranking be used?

(1) A database of dielectric constants made by calculation

Dielectric constants are a property for which few databases of measured values exist. Petousis and co-workers used a method called density functional perturbation theory (DFPT) to compute dielectric constants and refractive indices, and published a database of dielectric tensors for 1,056 compounds. Noting that the number of compounds with a known dielectric constant is a few hundred, the paper calls it the largest dielectric-tensor database to date. The data are integrated with the Materials Project and publicly available Sourced [Reference 7].

DFPT is a method that computes how electrons and atoms respond to a small external push such as an electric field (commentary). It is a way to fill in, by calculation, a property for which measurements are few.

(2) Comparison with experiment: value error and rank agreement

How far a calculated value can be trusted has been examined by earlier work. Petousis and co-workers compared the eigenvalues of the dielectric tensors from DFPT with literature measurements on single crystals and polycrystals for 88 compounds. The mean relative deviation (MARD) of GGA/PBE was 16.2%, smaller than for LDA, and that of the refractive index was 5.7%. They add, however, that accuracy is lowest for compounds with complex structural relaxation, such as perovskites with octahedral rotations, and for compounds with strong anharmonicity. On the other hand, the Spearman correlation coefficient, which expresses agreement in ranking, was high at 0.92, and the authors conclude that even with errors in the values, the method is useful for finding candidates by ranking Sourced [Reference 8].

We checked this "ranking survives the error" feel with synthetic data.

• Data (made up): 1,000 "compounds", each given a true value imitating a dielectric constant, spread log-uniformly between 3 and 200.
• Computed values: a random error was applied to the true values so that the mean relative deviation is about 16% (matched to the figure in Petousis et al.).
• Question: if we pick the top X% by computed value, what share of the true top 5% do we find?

Result (median over 200 random seeds): the rank correlation was 0.987, so the ranking was almost preserved. Choosing only the top 5% by computed value found just 66% of the true top 5%; widening to the top 10% found 94%, and the top 20% found 100% Our calculation.

Our calculation. Python (NumPy), double precision. The data are made up for this article and are not properties of real compounds. The rank correlation of the real data (0.92) is lower than that of the synthetic data (0.987) probably because the errors in real data are not random but concentrated in particular compounds, such as those with complex structural relaxation (commentary).

5. Thermoelectric materials: the distance from calculation to experiment

Thermoelectric materials are a field where the distance between the "expectation" and the "reality" of MI is often discussed. Here we read, in turn, the data, narrowing by calculation, cases verified by experiment, and the view of a recent review.

(1) How measured data were collected (Gaultois et al., 2013)

Gaultois and co-workers extracted information from more than 100 publications to build a large database of thermoelectric materials (more than 18,000 data points, several classes of compounds, values measured at several temperatures). They say that visualizing the data shows that a promising thermoelectric candidate needs to show an electrical resistivity close to 1 mΩ·cm at 300 K, that is, a conductivity far above the Mott minimum metallic conductivity. They also organize the supply risk of materials using an index of elemental reserves and production concentration (the Herfindahl-Hirschman index) Sourced [Reference 4].

How to read it: this is a database of collected measured values. Because the way samples were made, the measurement conditions and the temperature differ from paper to paper, work to make the data quality consistent is needed before machine learning (commentary).

(2) Narrowing by calculation: lattice thermal conductivity of half-Heusler compounds (Carrete et al., 2014)

Carrete and co-workers built a method that combines machine learning, physical insight and automatic first-principles calculations to estimate the lattice thermal conductivity (the contribution of lattice vibrations to heat transport) for many compounds. Scanning about 79,000 half-Heusler entries in AFLOWLIB.org, they found that among 450 mechanically stable ordered semiconductors the lattice thermal conductivity spans more than two orders of magnitude, and a screening of thermodynamic stability then narrowed the list to 77 systems. Among them they proposed three semiconductors with thermal conductivity below 5 W/(m·K) for experimental study Sourced [Reference 9].

A low thermal conductivity is one condition for raising zT (the formula in chapter 2). But this is narrowing by calculation, and within the abstract it does not include results of actually making and measuring the three proposed compounds (commentary; we did not check the full text).

(3) Cases confirmed by experiment

• A recommendation engine and RE12Co5Bi (Gaultois et al., 2016): they built an open machine-learning recommendation engine that pre-screens about 25,000 known materials and suggests promising compositions. In this example they actually synthesized and measured the RE12Co5Bi family (RE = Gd, Er), which the engine had predicted, a compound system with a high loading of metallic elements, and report that the predicted trends, low thermal conductivity, high electrical conductivity and a modest Seebeck coefficient, were confirmed experimentally. The paper presents this system as a candidate new thermoelectric material worth further study Sourced [Reference 12]. It does not say that the zT value is high.
• Cu_1-xAg_xGaTe₂ (Zhong et al., 2023): the authors report that p-type chalcogenides predicted by an AI approach and then actually synthesized reached a peak zT of about 1.90 at 770 K for x = 0.45 Sourced [Reference 13]. The value is the authors' measurement, and we could not find independent replication.

(4) The view of a review: "predictions are accurate, yet new materials are slow to appear"

The review by Athar and Jund (2026) says that machine learning reports impressive test scores, with R² of 0.90 to 0.98, for complex thermoelectric properties (zT, power factor, electrical and thermal conductivity), but that only a handful of predictions led to the experimental discovery of new high-zT materials. It names as main obstacles the small-data problem, sampling bias in cross-validation, inadequate structural representation and thermodynamic phase stability, and argues that standard random-split validation tends to overestimate performance because it ignores clusters (hidden hierarchies) among chemical families Sourced [Reference 17]. As remedies it proposes PCA-based sampling and an active-learning closed loop that links fast stability filters (such as GNoME) with high-throughput thin-film synthesis [Reference 17].

This is the view of a review, not a fact that holds for every study. But it points in the same direction as the "difference of columns" and the "pitfalls of splitting" seen in chapters 3 and 4. The pitfalls of splitting are covered in detail in our explainer on how to read evaluation and reproducibility in MI.

6. Two-dimensional materials: 1,036 "easily exfoliable" by calculation

2D materials are crystals only a few atomic layers thick; graphene is the typical example. The starting point is whether a thin sheet can be peeled from a layered crystal. Searches by calculation for compounds that "peel easily" have been carried out on a large scale.

(1) A computational search for ease of exfoliation (Mounet et al., 2018)

Starting from 108,423 experimentally known three-dimensional compounds, Mounet and co-workers selected 5,619 compounds that appear layered by geometric and bonding criteria. They computed binding energies with DFT including van der Waals forces (validated against experimental structural data and random-phase-approximation binding energies) and identified 1,825 compounds as easily or potentially exfoliable. Of these, 1,036 are easily exfoliable. The paper gives as background that only a few dozen 2D materials have so far been synthesized or exfoliated Sourced [Reference 11].

The C2DB (Computational 2D Materials Database) of Haastrup and co-workers computes structural, thermodynamic, elastic, electronic, magnetic and optical properties of about 1,500 2D materials with DFT and many-body perturbation theory (the Bethe-Salpeter equation as well for about 250 materials), by a semi-automated workflow that is consistent throughout, and makes them public. The paper says it identified "a large number of new potentially synthesisable 2D materials" Sourced [Reference 16].

(2) The difficulty of predicting exfoliation energy by machine learning

Matbench has a task that predicts the exfoliation energy (meV/atom) of 2D materials from crystal structure (636 records). The data are calculations from JARVIS-DFT (OptB88vdW and TBmBJ functionals), not measurements. On the official leaderboard (checked 22 October 2026), the best method (MODNet) has an MAE of 33.2 meV/atom, about half of the Dummy baseline (always answering the mean) of 67.3 meV/atom Sourced [References 1, 6]. It is an example of how large the improvement from machine learning is on a task with few records (commentary).

7. Organic LED materials and semiconductor process development

(1) A search for organic LED molecules: an example that links calculation, synthesis and devices (Gómez-Bombarelli et al., 2016)

Gómez-Bombarelli and co-workers reported the design of organic LED materials in which theory, quantum chemistry, cheminformatics, machine learning, industrial expertise, organic synthesis, molecular characterization, device fabrication and optoelectronic testing were linked in one flow. In a search space of 1.6 million molecules, they computed more than 400,000 with time-dependent DFT and found thousands of promising new organic LED molecules. The team selected the best candidates and synthesized them, and the external quantum efficiencies measured experimentally reached as high as 22% Sourced [Reference 14].

This example is one of the cases in this article in which the chain runs from calculation to synthesis and device measurement. "Found thousands" is a number of candidates from calculation; what was actually synthesized and measured is a subset chosen from them (from the description in the abstract; we did not check the full text). Note also that the process includes a step in which people (the team) pick the candidates.

(2) Semiconductor process development: combining humans and algorithms (Kanarik et al., 2023)

In semiconductor manufacturing, developing plasma process conditions is an expensive task. To see whether AI could lower this cost, Kanarik and co-workers compared Bayesian-optimization algorithms and human engineers in a controlled virtual process game. The result was that human engineers excel in the early stages, while the algorithms are far more cost-efficient near the tight tolerances of the target. With a "human first, computer last" strategy, the cost to reach the target was reduced to half of that for human designers alone Sourced [Reference 15]. The paper also notes cultural challenges in partnering humans with computers.

8. A materials engineer's view: connecting calculated values to experiments

Let us reorder the cases so far by "how far they have been verified". The marks are our organization based on the range of confirmation stated in the abstract of each paper.

Many cases end at calculation, and few cases "made and measured" the material. We found no case confirmed on a real device or in production within the scope of our survey. On this premise, here are three viewpoints for handling electronic-materials data and calculations in your own work (all are our commentary, not an official guideline).

9. Limits, what we could not confirm / summary

(1) Depth of checking of the papers

We checked the papers by Petousis and co-workers, Zhuo and co-workers, Gaultois and co-workers (2013 and 2016), Carrete and co-workers, Mounet and co-workers, Haastrup and co-workers, Gómez-Bombarelli and co-workers, Kanarik and co-workers and Zhong and co-workers within the abstract (or a published summary). We checked the text of Borlido and co-workers (2020) and of Dunn and co-workers (Matbench), and the official pages of the Materials Project and Matbench. Details not in the abstracts (the number of experiments, the split method, the basis of comparison) were not checked.

(2) Items for which experimental confirmation could not be confirmed

We could not confirm the results of actually measuring the three semiconductors proposed by Carrete and co-workers, how many of the 1,036 of Mounet and co-workers were verified by experiment, or the number of molecules synthesized by Gómez-Bombarelli and co-workers and the individual device performances.

(3) Independent replication and practical use

The zT of 1.90 of Zhong and co-workers, the external quantum efficiency of 22% of Gómez-Bombarelli and co-workers and the RE12Co5Bi results of Gaultois and co-workers are treated as claims of the papers. We could not confirm independent replication, nor adoption in real devices or mass production, from primary sources.

(4) What we did not cover

Solar cells (perovskites and others), inorganic light-emitting diode materials, power semiconductors (SiC, GaN), memory materials, magnetic materials and atomistic simulation of semiconductors with machine-learning interatomic potentials are not covered. Perovskite solar cells are covered in our explainer on perovskite basics and related articles, and machine-learning interatomic potentials in our explainer on machine-learning interatomic potentials.

(5) Comparing numbers

Data, calculation conditions, splits and metrics differ from paper to paper. This article does not rank any method, material or research group.

10. How to read the yardsticks and standards

The quality of MI for electronic materials is measured with three yardsticks: the gap between calculated and measured values, the prediction error of machine learning and the share of good materials missed when candidates are picked by ranking. The first shows "how close calculation is to measurement", the second "how well machine learning reproduces its training labels", and the third "whether narrowing loses truly good materials". No single one of them tells the whole story. Meanwhile, measurements have test-method standards (JIS, ISO, IEC). These define how to measure; they do not decide how good a calculation or a machine-learning model is. We could not find a standard specific to MI for electronic materials. The checking date is 22 October 2026.

The arrangement of the table is Our calculation (organization). Numbers, titles and publication dates of standards were checked on the official ISO pages and the JIS search of the Japanese Industrial Standards Committee (JISC) (checked 22 October 2026). The full texts are paid, so we stay with the scope statements. ASTM standards are not covered in this article, and we did not consult the ASTM site.

Cautions when reading (for materials engineers)

• A leaderboard score is not the strength of your own material: Matbench scores are for specific data (DFT calculations or collected measurements) under a fixed split. It is worth re-measuring on your own material systems and your own measurement conditions.
• Prediction and experimental verification are different questions: even with a good R² or MAE, it remains a "computed prediction" until you make and measure the material. Conversely, one hit in an experiment must be checked for chance with repeats and a control.
• Data leakage and the pitfall of splitting: if materials of the same chemical system or prototype fall on both the training and test sides, scores are inflated. The review points out that random-split validation leads to overestimation [17].
• Read standards and principles as recommended or required: JIS and ISO standards define measurement methods and terms, and are generally understood to become binding only when cited in laws or contracts (commentary). FAIR is a set of principles (recommended).

11. Glossary

12. References (primary sources)

1. Dunn A., Wang Q., Ganose A., Dopp D., Jain A. "Benchmarking materials property prediction methods: the Matbench test set and Automatminer reference algorithm", npj Computational Materials 6, 138 (2020), DOI 10.1038/s41524-020-00406-3 (open access, full text checked) - doi.org/10.1038/s41524-020-00406-3
2. Borlido P., Aull T., Huran A. W., Tran F., Marques M. A. L., Botti S. "Large-Scale Benchmark of Exchange-Correlation Functionals for the Determination of Electronic Band Gaps of Solids", Journal of Chemical Theory and Computation 15, 5069-5079 (2019), DOI 10.1021/acs.jctc.9b00322 (abstract checked) - doi.org/10.1021/acs.jctc.9b00322
3. Borlido P., Schmidt J., Huran A. W., Tran F., Marques M. A. L., Botti S. "Exchange-correlation functionals for band gaps of solids: benchmark, reparametrization and machine learning", npj Computational Materials 6, 96 (2020), DOI 10.1038/s41524-020-00360-0 (open access, full text checked) - doi.org/10.1038/s41524-020-00360-0
4. Gaultois M. W., Sparks T. D., Borg C. K. H., Seshadri R., Bonificio W. D., Clarke D. R. "Data-Driven Review of Thermoelectric Materials: Performance and Resource Considerations", Chemistry of Materials 25, 2911-2920 (2013), DOI 10.1021/cm400893e (abstract checked) - doi.org/10.1021/cm400893e
5. Materials Project documentation "Electronic Structure" (statements on band-gap accuracy, checked 22 October 2026) - docs.materialsproject.org/methodology/materials-methodology/electronic-structure
6. Matbench official pages (leaderboard and data description for each task, checked 22 October 2026) - matbench.materialsproject.org
7. Petousis I., Mrdjenovich D., Ballouz E., Liu M., Winston D., Chen W., Graf T., Schladt T. D., Persson K. A., Prinz F. B. "High-throughput screening of inorganic compounds for the discovery of novel dielectric and optical materials", Scientific Data 4, 160134 (2017), DOI 10.1038/sdata.2016.134 (open access, abstract checked) - doi.org/10.1038/sdata.2016.134
8. Petousis I., Chen W., Hautier G., Graf T., Schladt T. D., Persson K. A., Prinz F. B. "Benchmarking density functional perturbation theory to enable high-throughput screening of materials for dielectric constant and refractive index", Physical Review B 93, 115151 (2016), DOI 10.1103/PhysRevB.93.115151 (abstract checked) - doi.org/10.1103/physrevb.93.115151
9. Carrete J., Li W., Mingo N., Wang S., Curtarolo S. "Finding Unprecedentedly Low-Thermal-Conductivity Half-Heusler Semiconductors via High-Throughput Materials Modeling", Physical Review X 4, 011019 (2014), DOI 10.1103/PhysRevX.4.011019 (abstract checked) - doi.org/10.1103/physrevx.4.011019
10. Zhuo Y., Mansouri Tehrani A., Brgoch J. "Predicting the Band Gaps of Inorganic Solids by Machine Learning", The Journal of Physical Chemistry Letters 9, 1668-1673 (2018), DOI 10.1021/acs.jpclett.8b00124 (abstract checked) - doi.org/10.1021/acs.jpclett.8b00124
11. Mounet N., Gibertini M., Schwaller P., Campi D. et al. "Two-dimensional materials from high-throughput computational exfoliation of experimentally known compounds", Nature Nanotechnology 13, 246-252 (2018), DOI 10.1038/s41565-017-0035-5 (abstract checked) - doi.org/10.1038/s41565-017-0035-5
12. Gaultois M. W., Oliynyk A. O., Mar A., Sparks T. D., Mulholland G. J., Meredig B. "Perspective: Web-based machine learning models for real-time screening of thermoelectric materials properties", APL Materials 4, 053213 (2016), DOI 10.1063/1.4952607 (abstract checked) - doi.org/10.1063/1.4952607
13. Zhong Y., Hu X., Sarker D., Su X., Xia Q., Xu L., Yang C., Tang X., Levchenko S. V., Han Z., Cui J. "Data analytics accelerates the experimental discovery of Cu_1-xAg_xGaTe₂ based thermoelectric chalcogenides with high figure of merit", Journal of Materials Chemistry A 11, 18651-18659 (2023), DOI 10.1039/D3TA03990K (abstract checked) - doi.org/10.1039/D3TA03990K
14. Gómez-Bombarelli R., Aguilera-Iparraguirre J., Hirzel T. D. et al. (21 authors) "Design of efficient molecular organic light-emitting diodes by a high-throughput virtual screening and experimental approach", Nature Materials 15, 1120-1127 (2016), DOI 10.1038/nmat4717 (abstract checked) - doi.org/10.1038/nmat4717
15. Kanarik K. J., Osowiecki W. T., Lu Y., Talukder D. et al. "Human-machine collaboration for improving semiconductor process development", Nature 616, 707-711 (2023), DOI 10.1038/s41586-023-05773-7 (abstract checked) - doi.org/10.1038/s41586-023-05773-7
16. Haastrup S., Strange M., Pandey M., Deilmann T. et al. "The Computational 2D Materials Database: high-throughput modeling and discovery of atomically thin crystals", 2D Materials 5, 042002 (2018), DOI 10.1088/2053-1583/aacfc1 (abstract checked) - doi.org/10.1088/2053-1583/aacfc1
17. Athar S., Jund P. "Beyond predicted zT: Machine learning strategies for the experimental discovery of thermoelectric materials", Artificial Intelligence Chemistry 4, 100113 (2026), DOI 10.1016/j.aichem.2026.100113 (abstract checked on arXiv:2601.06571) - doi.org/10.1016/j.aichem.2026.100113
18. Wilkinson M. D. et al. "The FAIR Guiding Principles for scientific data management and stewardship", Scientific Data 3, 160018 (2016), DOI 10.1038/sdata.2016.18 - doi.org/10.1038/sdata.2016.18
19. Japanese Industrial Standards Committee (JISC) JIS search (numbers and titles of JIS C 2139-2-1, C 2139-2-2, R 1650-1, R 1650-2 and R 1650-3 checked; titles in Japanese) - jisc.go.jp/app/jis/general/GnrJISSearch.html
20. 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
21. ISO "ISO/TS 80004-13:2024 Nanotechnologies - Vocabulary - Part 13: Graphene and other two-dimensional (2D) materials" (ISO page) - iso.org/standard/82855.html
22. ISO/IEC "ISO/IEC 22989:2022 Information technology - Artificial intelligence - Artificial intelligence concepts and terminology" (ISO page) - iso.org/standard/74296.html

13. Claim-to-source audit

Last updated: 22 October 2026 / Sources are limited to primary ones (peer-reviewed papers, arXiv preprints, the official Materials Project and Matbench pages, and the official ISO and JISC pages). The full texts of ISO standards and JIS standards were not checked. The synthetic-data example is a calculation on data made up for this article, not properties of real compounds. Experimental confirmation of the proposed candidates, independent replication, adoption in real devices or production, some details in the text of certain papers, and a standard specific to MI for electronic materials could not be confirmed and are therefore not stated. The figures are explanatory concept diagrams, drawings that include our own calculation, or plots of published numbers; Figs. 1 and 9 are ImageGen-generated conceptual illustrations.