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

• A descriptor is a “business card” of a material in numbers. It turns composition, structure and process into a list of values a machine can read. The choice of items on the card changes how well predictions come true
• Data quality decides prediction quality. Measurement scatter, differences in scale, missing values, bias in the data and whether failures are recorded all distort the result. Tidying the data is worth more than cleverness in the model
• The one point to remember: “could not be measured” and “did not work” are also data. If you keep them with a reason instead of leaving a blank, the “boundary” of where MI can be used becomes visible

2. What a descriptor is: a material as a “business card” in numbers

(1) An analogy: comparing people you have just met, by their business cards

When you meet someone new, a business card helps you compare people. It has fixed items in a row: company, title, location. Materials work the same way. If there is a business card with fixed items written as numbers, a machine can calculate how similar two materials are. That card is a descriptor (also called a fingerprint). The review by Ramprasad and colleagues explains that predictions take the form “first express the material as numbers, then link that representation to the target property through a learned mapping,” and that these numerical representations can be of many types and scales depending on the applicationSourced [Ref. 1].

View original SVG, values and sources

Original SVG reference: Chart including our calculation (vector drawing). The idea of a "business card": a material turned into numbers from its composition, structure and process. Only the composition rows are actual calculations; the others show example items.

The items on a card fall roughly into three groups, according to where they come from: the composition (what is in it, and in what proportions), the structure (how the atoms are arranged), and the process (temperature, time, atmosphere and the like). As in Fig. 2, composition items can be calculated from the chemical formula alone. Structure and process items need other information.

The arrangement of this table is our own commentary. For other representations (spectra, images, text) and a detailed comparison of structure-based methods, see the reviews by Musil et al. [Ref. 9] and Damewood et al. [Ref. 10].

3. Making a card from composition: checking by hand

Let us build the simplest card by hand. Our example material is LiFePO₄ (lithium iron phosphate). It contains lithium (Li), iron (Fe), phosphorus (P) and oxygen (O) in the ratio 1 : 1 : 1 : 4, seven atoms in all.

1. Find the fractions. Li, Fe and P are 1/7 each, and O is 4/7
2. Look up an element property. We use electronegativity (a rough measure of how strongly an atom pulls electrons). The PubChem periodic-table values are Li 0.98, Fe 1.83, P 2.19 and O 3.44 [Ref. 19]
3. Take the fraction-weighted mean. (0.98 + 1.83 + 2.19) ÷ 7 + 4 × 3.44 ÷ 7 = 0.714 + 1.966 = 2.68
4. Take the range. maximum − minimum = 3.44 − 0.98 = 2.46

The table at the bottom of Fig. 3 shows the “cards” made by the same steps for four compounds (Our calculation). The mean electronegativity is 2.68 for LiFePO₄, 2.55 for BaTiO₃, 2.05 for NaCl and 2.38 for MgO. The range (the difference between largest and smallest) can be read as larger for compounds with stronger ionic character (commentary). The “elemental property statistics” in the framework of Ward et al. extend this idea (mean, range and so on) to 22 element properties [Ref. 3].

Our calculation. Element values are from PubChem's periodic-table data [Ref. 19]. Means are weighted by atom fractions. This is a two-item introductory example; real composition descriptors have dozens to hundreds of items [Ref. 3].

4. Putting lab notes into a table: one row per sample, and data quality

When you use your own experimental data, the first job is to put the laboratory notebook into a table. Machines read tables. Text and photos alone cannot be turned into numbers. Fig. 4 shows an example of turning free-text notes into a table with one row per sample.

View original SVG, values and sources

Original SVG reference: Conceptual diagram (vector drawing). An example of turning notes into a machine-readable table: conditions and results in separate columns, with units, lots and failures recorded. Values are for illustration.

There are only three rules for the table. (1) One row per sample (if you measure the same sample several times, add that many rows). (2) One column per item (put the unit in the column name). (3) Do not leave blanks (if something could not be measured or failed, write why). Just this makes later work much easier.

This table is our own commentary. The international standards on data quality (the ISO/IEC 5259 series) are introduced by title and scope in Chapter 11.

Here is one example of measurement scatter in numbers. The OQMD paper reports that the mean absolute error between experiments measuring the formation energy of the same compound is 0.082 eV/atom, “surprisingly large” (the DFT-versus-experiment difference is 0.096 eV/atom)Sourced [Ref. 16]. How to handle measurement scatter is organized in the field of metrology. For example, JCGM 100:2008 (the GUM) defines uncertainty of measurement as a parameter, associated with the result of a measurement, that characterizes the dispersion of the values that could reasonably be attributed to the measurandSourced [Ref. 26].

5. Putting numbers on the same scale

The items on a card often differ by orders of magnitude. Temperature, for example, runs from 300 to 1,300 K (a span of about 1,000), while a composition fraction runs from 0 to 1 (a span of 1). Many machine-learning methods are affected by such differences. The scikit-learn user guide explains that many estimators may behave badly if the individual features do not more or less look like standard normally distributed data (Gaussian with zero mean and unit variance)Sourced [Ref. 17].

We check the effect with a small experiment (Our calculation). We use a nearest-neighbor method, which “returns the mean value of the three closest past samples,” with two descriptors, temperature and fraction. The true property depends strongly (and periodically) on the fraction and only gently on temperature. We train on 60 points and measure the error on 500 new points (average of 30 runs).

The result is in Fig. 6. Without putting the numbers on one scale, distance is decided almost entirely by the difference in temperature, and the effect of the fraction is not seen. The error (RMSE) is 0.83, worse than the do-nothing prediction that just returns the mean (0.73). After standardization (subtracting each item's mean and dividing by its spread), the error falls to 0.24.

Our calculation. Synthetic data (true property = sin(2π × fraction) + 0.3 × (temperature − 800)/500 + noise 0.05). Not a result on real materials. For standardization, the mean and spread were computed from the training data only and applied to the test data (to avoid data leakage).

6. Missing values and “could not be measured”

Blanks (missing values) are common in materials tables: items not measured, samples that could not be measured, values outside the instrument's range. The scikit-learn user guide explains that a basic strategy for using incomplete datasets is to discard whole rows or columns containing missing values, but this comes at the price of losing data that may be valuable; a better strategy is to impute the missing values, that is, to infer them from the known part of the data [Ref. 18].

But why the values are missing matters. If values are not missing by chance, for example when “the lower the value, the more likely it falls below the detection limit and cannot be measured,” then collecting only the measured points creates a bias. We check this with a small experiment (Our calculation).

• Setup. Descriptor x (for example, the amount of an additive) and property y are related by y = x + noise (standard deviation 0.5). Values of y below −0.5 fall under the detection limit and are not recorded (about 33% of all values are missing)
• The true mean is 0.00. But the mean of the measured points alone rises to 0.60
• The slope is 1.00 for all points, but 0.82, gentler, for the measured points alone
• If we fill the missing values with the detection-limit value (−0.5), the mean is 0.24, still different from the true 0.00

Our calculation. Synthetic data (means and slopes computed from 200,000 points; the plot shows 200 points). Not a result on real materials.

7. Biased data, and the danger of extrapolation

If the data you collected are concentrated in part of the range of conditions, predictions will not come true outside that range. We check this with a small experiment (Our calculation). Let the true relation be y = sin(3x) + 0.5x, and train on 30 points (noise 0.05) measured only in x = 0 to 1. We evaluate inside that range and at x = 1 to 3 (outside).

The result is in Fig. 8. Inside the range, the error is 0.03 for the cubic polynomial and 0.08 for the 3-nearest-neighbor method; both are accurate. Outside (x = 1 to 3), the polynomial's error jumps to 12.6 (median). The 3-nearest method stays constant at the edge value, with an error of 0.98. That is about the same as the spread of the true values outside (standard deviation 0.97), no better than a “return the mean” prediction.

Our calculation. Synthetic data. Average of 50 runs (median for the polynomial). Outside the range, the way models fail depends on the model type (running away, or going flat), but neither is right.

This is an artificial and extreme example. But similar bias has been reported in real materials data. Jia and colleagues found, in reports of hydrothermal synthesis of amine-templated metal oxides, that the choice of amine follows a power-law distribution in which 17% of the amines occur in 79% of the reported compounds. Fashion and cognitive bias in how people choose experiments show up in the data. They further showed with 548 randomly generated experiments that the popularity of reactants and the choice of reaction conditions are uncorrelated with whether the reaction succeeds, and reported that models trained on a smaller randomized dataset outperformed models trained on larger human-selected datasetsSourced [Ref. 13].

Li and colleagues also showed, for several large materials datasets, that up to 95% of the data can be removed from training with little effect on in-distribution prediction performance, that the removable data are tied to over-represented material types, and that the degradation out of distribution is not preventedSourced [Ref. 15]. What matters more than the amount of data is what range it covers, and how evenly.

8. The value of failed results

The data that appear in papers and reports are mostly successful results. Failed conditions are often not written down. But records of failure show the “boundary” of the conditions where success happens. The left side of Fig. 9 is the case where only successes are recorded: the successful points gather around popular conditions, and you cannot tell how far the conditions can be widened. The right side is the case where failures are recorded too, and the boundary between success and failure shows.

Raccuglia and colleagues collected data on “failed reactions” left in laboratory notebooks (reactions that never became papers) and had a machine-learning model learn whether reactions of templated vanadium selenites succeed. The model was reported to predict reaction success and failure more accurately than human intuitionSourced [Ref. 12]. The description of the paper also notes that failed reactions carry valuable information about the limits of the reaction conditions, yet are rarely reported [Ref. 12].

Kononova and colleagues used text-processing techniques to extract solid-state synthesis “recipes” (target material, starting compounds, operations with their conditions, and the balanced reaction) from papers, and assembled 19,488 synthesis entries (from 53,538 synthesis paragraphs) into a public datasetSourced [Ref. 14]. It is one example of turning text into tables. The bias of the original papers (Chapter 7), however, is carried over as it is (commentary).

9. How to choose and reduce descriptors

More card items are not always better. Check these four points (commentary).

• Use only information available at prediction time. Do not put on the card information that can be known only after the result is known (for example, another property decided after measurement). The evaluation number looks better, but the model is useless (a kind of data leakage)
• Does it have physical or chemical meaning? Meaningful items work more easily with little data, and the result can be interpreted. Ghiringhelli et al. discuss the requirements for a suitable descriptor and, for a classic example (the energy difference between zinc blende or wurtzite and rocksalt semiconductors), show how a meaningful descriptor can be found systematically [Ref. 2]
• Is the number of items in proportion to the number of data? When the number of items approaches the number of samples, overfitting becomes likely (see our explainer on machine learning with small data). SISSO by Ouyang et al. is a compressed-sensing method that selects a small set of best descriptors from a huge number of candidates, and is reported to give stable results even with small training setsSourced [Ref. 11]
• Can it be measured or computed? If an item on the card cannot be obtained cheaply and reliably in the field, it cannot be used

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

(1) What we could not confirm

• The electronegativity scale in PubChem. We could not confirm from PubChem's periodic-table page which scale (Pauling or another) the values use. So this article writes “PubChem value” and does not name a scale
• Which descriptor is best. We do not cover which descriptor works best for which material and property. Comparisons change with dataset and split (see the reviews by Musil et al. [Ref. 9] and Damewood et al. [Ref. 10])
• Numerical experiments. The calculations in Chapters 5 to 7 use synthetic data. They do not show how large the effects of bias and missing values are in real materials
• Decision-tree-based methods and scale. We did not check primary sources on whether some methods are insensitive to differences in scale
• Text of standards. We did not confirm the requirements in the text of ISO/IEC 5259, ISO 5725-1 and JCGM 100 (we only quote titles, publication dates and a definition)
• An MI-specific data-quality standard. We could not confirm one (Chapter 11)

11. How to read the yardsticks and standards

The quality of descriptors and data cannot be measured directly. Usually we measure it indirectly, through the performance of a model trained with that card and those data (its error and how the data were split). The yardsticks are read as follows. (1) Compare the error (MAE, RMSE) with the measurement scatter (reproducibility). (2) The numbers change with how the data are split (randomly, or by composition or lot). (3) Benchmarks compare methods on common data. (4) Data quality has its ideas organized in international standards on data quality and standards on uncertainty of measurement. The table lists what we could confirm (checked 22 October 2026). We could not confirm a standard specific to descriptors or data quality in MI.

The arrangement of the table is Our calculation (organization). ISO/IEC and ISO 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 model's number cannot exceed the quality of the data. With data of large scatter, missing values and bias, any model has limits. Read the error against the measurement scatter
• A benchmark ranking differs from performance on your material. Results on public data such as Matbench were obtained under conditions different from your card and your data
• Split traps. If repeats of the same sample or lot fall in both training and evaluation, the number is overstated. Do preprocessing (standardizing, imputing, choosing items) on the training data only
• Most principles and standards are recommendations. The FAIR principles are guidance without legal force. ISO/IEC standards generally bind only when cited in law or contracts (commentary). When something says “compliant,” check which requirement of which document is meant

12. Glossary

13. References (primary sources)

1. 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
2. Ghiringhelli L. M. et al. "Big Data of Materials Science: Critical Role of the Descriptor", Phys. Rev. Lett. 114, 105503 (2015), DOI 10.1103/PhysRevLett.114.105503 — doi.org/10.1103/PhysRevLett.114.105503
3. Ward L. et al. "A general-purpose machine learning framework for predicting properties of inorganic materials", npj Comput. Mater. 2, 16028 (2016), DOI 10.1038/npjcompumats.2016.28 (open access) — doi.org/10.1038/npjcompumats.2016.28
4. Ward L. et al. "Matminer: An open source toolkit for materials data mining", Comput. Mater. Sci. 152, 60–69 (2018), DOI 10.1016/j.commatsci.2018.05.018 — doi.org/10.1016/j.commatsci.2018.05.018
5. Rupp M. et al. "Fast and Accurate Modeling of Molecular Atomization Energies with Machine Learning", Phys. Rev. Lett. 108, 058301 (2012), DOI 10.1103/PhysRevLett.108.058301 — doi.org/10.1103/PhysRevLett.108.058301
6. Bartók A. P., Kondor R., Csányi G. "On representing chemical environments", Phys. Rev. B 87, 184115 (2013), DOI 10.1103/PhysRevB.87.184115 — doi.org/10.1103/PhysRevB.87.184115
7. Behler J., Parrinello M. "Generalized Neural-Network Representation of High-Dimensional Potential-Energy Surfaces", Phys. Rev. Lett. 98, 146401 (2007), DOI 10.1103/PhysRevLett.98.146401 — doi.org/10.1103/PhysRevLett.98.146401
8. Rogers D., Hahn M. "Extended-Connectivity Fingerprints", J. Chem. Inf. Model. 50, 742–754 (2010), DOI 10.1021/ci100050t — doi.org/10.1021/ci100050t
9. Musil F. et al. "Physics-Inspired Structural Representations for Molecules and Materials", Chem. Rev. 121, 9759–9815 (2021), DOI 10.1021/acs.chemrev.1c00021 — doi.org/10.1021/acs.chemrev.1c00021
10. Damewood J. et al. "Representations of Materials for Machine Learning", Annu. Rev. Mater. Res. 53, 399–426 (2023), DOI 10.1146/annurev-matsci-080921-085947 — doi.org/10.1146/annurev-matsci-080921-085947
11. Ouyang R. et al. "SISSO: A compressed-sensing method for identifying the best low-dimensional descriptor in an immensity of offered candidates", Phys. Rev. Mater. 2, 083802 (2018), DOI 10.1103/PhysRevMaterials.2.083802 — doi.org/10.1103/PhysRevMaterials.2.083802
12. 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
13. Jia X. et al. "Anthropogenic biases in chemical reaction data hinder exploratory inorganic synthesis", Nature 573, 251–255 (2019), DOI 10.1038/s41586-019-1540-5 — doi.org/10.1038/s41586-019-1540-5
14. Kononova O. et al. "Text-mined dataset of inorganic materials synthesis recipes", Sci. Data 6, 203 (2019), DOI 10.1038/s41597-019-0224-1 (open access) — doi.org/10.1038/s41597-019-0224-1
15. 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
16. 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
17. scikit-learn User Guide "Preprocessing data" — scikit-learn.org/stable/modules/preprocessing.html
18. scikit-learn User Guide "Imputation of missing values" — scikit-learn.org/stable/modules/impute.html
19. PubChem (NIH National Library of Medicine) "Periodic Table of Elements" (table of element properties; this article used the element data from the REST API) — pubchem.ncbi.nlm.nih.gov/periodic-table/
20. 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
21. 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
22. Kapoor S., Narayanan A. "Leakage and the reproducibility crisis in machine-learning-based science", Patterns 4, 100804 (2023), DOI 10.1016/j.patter.2023.100804 (open access) — doi.org/10.1016/j.patter.2023.100804
23. 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
24. 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
25. ISO "ISO 5725-1:2023 Accuracy (trueness and precision) of measurement methods and results — Part 1: General principles and definitions" (ISO page) — iso.org/standard/69418.html
26. JCGM "JCGM 100:2008 Evaluation of measurement data — Guide to the expression of uncertainty in measurement (GUM)" (PDF published by BIPM) — bipm.org/documents/…/JCGM_100_2008_E.pdf
27. 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

14. Claim-to-source audit

Last updated: 22 October 2026 / Sources are limited to primary materials (peer-reviewed papers and official pages of scikit-learn, PubChem, ISO and BIPM). The full text of ISO standards is paywalled and was not checked. The numerical experiments in Chapters 5 to 7 are calculations on synthetic data we generated, not results on real materials. We could not confirm the electronegativity scale in PubChem, which descriptor is best, decision-tree methods and scale, or an MI-specific data-quality standard, so these are not stated. All figures are explanatory concept diagrams or drawings that include our own calculation; the hero image and Fig. 5 are ImageGen-generated conceptual illustrations.