Research gapEngineering & technology

Materials science future directions: from predicted materials to ones we can actually make

Machine learning now proposes new inorganic materials faster than any laboratory can test them (Merchant et al., 2023; Zeni et al., 2025). The open question for the next decade is which of those predictions can be made, verified and scaled, and at what environmental cost (Park et al., 2026; Woods-Robinson & Trewartha, 2026).

PhD Research Lab InsightsEditorial team, reviewed by a subject specialist
Posted
19 July 2021
Last checked
7 October 2026
Reading time
10 min
Sources
10
Research proposal support
The gap

We cannot yet say reliably which computationally predicted inorganic materials can be synthesised, in what ordered or disordered form, by which route and with what environmental footprint. Discovery claims have outpaced verified synthesis and characterisation (Leeman et al., 2024; Park et al., 2026), and sustainability is rarely assessed at the early stage where it could still change the design (Woods-Robinson & Trewartha, 2026).

Methodological gap (synthesisability prediction)Evidence gap (experimental validation)Empirical gap (scale-up and sustainability)Practical-knowledge gap (automated characterisation)

What we know

  • Deep learning with active learning identified 2.2 million structures below the previous convex hull, of which 381,000 are newly stable on the updated hull (Merchant et al., 2023).
  • Generative models can now propose stable, new structures with target properties; one generated oxide was made and its measured bulk modulus came within about 20% of the target (Zeni et al., 2025).
  • An autonomous laboratory realised 36 compounds from 57 targets over 17 days of operation (Szymanski et al., 2023); a 2026 correction clarified that "new" meant new to the prediction platform, not necessarily new to science (Szymanski et al., 2026).
  • Re-analysis judged that about two thirds of the claimed successes were probably known compounds with compositional disorder, and that automated Rietveld analysis is not yet reliable (Leeman et al., 2024).
  • Low prediction error does not guarantee good discovery: models can label many unstable materials as stable near the decision boundary (Riebesell et al., 2025).

What we do not know yet

  • Which measures of synthesisability (thermodynamic, heuristic or learned) best predict that a material can be made (Park et al., 2026).
  • How to predict and characterise disorder, so that a predicted ordered crystal is not confused with a known disordered phase.
  • How often predicted materials keep their properties when made at scale and built into a device.
  • How to estimate energy, supply-chain and end-of-life impacts at the discovery stage, when data are incomplete (Woods-Robinson & Trewartha, 2026).
  • What standards of evidence a discovery claim should meet before it is called new.

Research questions a PhD could ask

  1. For a chosen materials family, how well do current synthesisability models predict which candidates a laboratory can make phase-pure?
  2. Can machine-learning analysis of powder diffraction distinguish ordered predicted structures from disordered known phases as reliably as an expert?
  3. How much does an early, prospective life-cycle screen change the ranking of candidate materials for one application?
  4. What closed-loop design, mixing autonomous experiments with human checks, gives the most verified discoveries per experiment?
  5. Which reporting checklist would let readers judge novelty, purity and reproducibility of an AI-assisted discovery?
Methods that could close the gap
ApproachWhy it fits
Benchmark studyTests synthesisability and stability models against held-out experimental outcomes, using discovery-relevant metrics rather than average error (Riebesell et al., 2025).
Closed-loop experimentationCouples prediction with automated synthesis, as in self-driving laboratories (Tom et al., 2024; Szymanski et al., 2023), but with expert review of every claimed product.
Characterisation method developmentBuilds and validates automated diffraction analysis that models site disorder, the weakness identified by Leeman et al. (2024).
Prospective life-cycle assessmentScreens candidates for energy, supply and end-of-life burdens at low technology readiness (Woods-Robinson & Trewartha, 2026).
Systematic review of discovery claimsAudits published AI-assisted discoveries for evidence of novelty, purity and reproduction (Cheetham & Seshadri, 2024; Leeman et al., 2024).
Gap typeIndustry and research both working on it

The problem

Computational workflows can propose millions of hypothetical compounds, but there is a wide divide between what can be calculated and what can be synthesised (Park et al., 2026; Merchant et al., 2023).

Why it happens

Computed thermodynamic stability is not the same as a new, functional material (Cheetham & Seshadri, 2024). Predictions often ignore disorder and synthesis conditions, and automated analysis of products is not yet reliable (Leeman et al., 2024). Sustainability is usually assessed only after composition and route are fixed (Woods-Robinson & Trewartha, 2026).

Where it is felt most

The gap is felt most in AI-led discovery of inorganic functional materials, where industry laboratories and universities are publishing large prediction sets and autonomous synthesis platforms (Merchant et al., 2023; Zeni et al., 2025; Szymanski et al., 2023), and in any sector that must scale a new material into a product (Woods-Robinson & Trewartha, 2026).

How big is the problem

  • 2.2 millionpredicted structures below the previous convex hull, with 381,000 newly stable materials, in one deep-learning study (Merchant et al., 2023).
  • 736of those predicted structures already independently realised by experiment at publication (Merchant et al., 2023).
  • 36 of 57targets realised by an autonomous laboratory in 17 days, as given after the 2026 correction (Szymanski et al., 2023; Szymanski et al., 2026).
  • About 2 in 3claimed autonomous-synthesis successes judged likely to be known disordered compounds on re-analysis (Leeman et al., 2024).
What studies say, and at what level of evidence
LevelWhat it shows
Lab and simulation studiesLarge-scale prediction and generative design now produce stable candidates at scale (Merchant et al., 2023; Zeni et al., 2025).
Trials or field studiesAutonomous synthesis has been demonstrated, but its novelty claims were corrected (Szymanski et al., 2023; Szymanski et al., 2026).
Critical re-analysesIndependent experts questioned novelty, credibility and characterisation of headline results (Cheetham & Seshadri, 2024; Leeman et al., 2024).
Systematic reviews and benchmarksReviews map self-driving laboratories (Tom et al., 2024); benchmarks show why task-relevant metrics matter (Riebesell et al., 2025).
PerspectivesProposals combine thermodynamics, chemical heuristics and machine learning to rank synthesisability (Park et al., 2026), and call for life-cycle thinking at the discovery stage (Woods-Robinson & Trewartha, 2026).
What is missingProspective, independently checked tests of synthesisability predictions, and discovery studies that report purity, disorder and environmental cost together.

Who should take this on

Scholar profile

  • A solid-state chemist with access to synthesis furnaces and powder diffraction.
  • A computational materials scientist who wants to test predictions against experiment.
  • A data scientist interested in automated characterisation and laboratory automation.
  • An industrial ecologist or engineer able to run prospective life-cycle assessment.

Subjects and PhD specialisations

  • Materials science and engineering
  • Solid-state and inorganic chemistry
  • Computational materials science
  • Machine learning for science
  • Chemical engineering (scale-up)
  • Industrial ecology and life-cycle assessment

Countries where the problem is greatest

  • United States: home of the A-Lab autonomous laboratory, a leading test bed for prediction-led synthesis (Szymanski et al., 2023).
  • United Kingdom: centre of independent re-analysis of autonomous-synthesis claims and of work on synthesisability metrics (Leeman et al., 2024; Park et al., 2026).
  • Countries building national self-driving laboratory or materials-acceleration programmes, which need validated workflows (Tom et al., 2024).

Scope for future research

  1. Synthesisability models validated prospectively in the laboratory, not only on historical databases (Park et al., 2026).
  2. Disorder-aware structure prediction and automated diffraction analysis (Leeman et al., 2024).
  3. Self-driving laboratories with built-in expert verification and open data (Tom et al., 2024).
  4. Early-stage life-cycle and supply-chain screening of candidate materials (Woods-Robinson & Trewartha, 2026).
  5. Shared reporting standards for AI-assisted discovery claims (Cheetham & Seshadri, 2024).

How a doctoral project could address it

  1. Choose one materials family and application

    For example, oxide electrocatalysts, lead-free piezoceramics or antibacterial alloys, with one measurable property.

  2. Review and audit

    Map published predictions and claimed syntheses in that family, noting how novelty and purity were shown (Leeman et al., 2024).

  3. Build the prediction set

    Rank candidates with open models and at least one synthesisability measure (Park et al., 2026).

  4. Make and characterise

    Synthesise a manageable number of candidates, with diffraction refined by hand as well as automatically.

  5. Add a prospective life-cycle screen

    Estimate energy and supply-chain burden for the shortlist (Woods-Robinson & Trewartha, 2026).

  6. Plan safety and data sharing

    Agree laboratory safety, data management and deposit of raw diffraction data.

  7. Report fully

    Publish negative results as well as successes, with a clear statement of what counts as new.

Why it is worth solving

Closing the gap between prediction and synthesis would turn very large computational databases into materials that industry can use, and would stop laboratory effort being spent on candidates that cannot be made (Park et al., 2026; Merchant et al., 2023). Clear evidence standards would also protect the credibility of AI-assisted discovery after high-profile corrections (Szymanski et al., 2026; Leeman et al., 2024).

Building sustainability into the first screen means that energy, supply and end-of-life burdens are considered while the design can still change, rather than after a material is locked into products (Woods-Robinson & Trewartha, 2026).

What has changed since 2021

Since then the biggest shift has come from computation. Deep learning has expanded the number of predicted stable inorganic crystals by an order of magnitude (Merchant et al., 2023), generative models can design candidates for a target property (Zeni et al., 2025), and autonomous laboratories plan and run syntheses with little human input (Szymanski et al., 2023; Tom et al., 2024).

The same period showed the limits. Experts asked how many predicted materials are new, credible or useful (Cheetham & Seshadri, 2024); a re-analysis of autonomous-synthesis results found that many claimed new compounds were probably known disordered phases (Leeman et al., 2024); and the original paper was corrected in 2026 (Szymanski et al., 2026). Making and proving a material is now the bottleneck (Park et al., 2026).

How the gap runs through the classic materials areas

The synthesis gap is not limited to one class of material. Each of the areas covered in the 2021 survey can host a doctoral project built around it. The questions below are suggestions for scoping, not findings.

Illustrative doctoral questions by materials area
AreaA synthesis-gap question a doctorate could ask
Energy oxides (perovskites)Can predicted oxide compositions be made phase-pure by a scalable route, and does disorder change performance?
Ceramics and piezoelectricsDo predicted lead-free compositions keep their properties after sintering at realistic temperatures?
Metals and alloysDo predicted antibacterial or implant alloys form the expected phases under industrial processing?
NanomaterialsHow reliably can automated characterisation confirm size, phase and purity?
Polymers and bio-based materialsCan prospective life-cycle screening rank candidate monomers before synthesis?

Is this feasible for a doctorate?

Yes, if the scope is narrow. A strong thesis might audit published claims in one materials family, test a synthesisability model on a small set of candidates made in the laboratory, and add a life-cycle screen. Access to synthesis equipment and diffraction, and a supervisory team that spans computation and experiment, matter more than access to the largest models. Negative results, such as predicted materials that could not be made, are a valid and useful contribution when reported carefully (Park et al., 2026).

Practical planning for a doctoral project on the synthesis gap
ConsiderationWhat to plan for
FacilitiesFurnaces or other synthesis routes, powder diffraction and time on shared instruments.
SkillsStructure refinement, scripting and basic machine learning across the supervisory team.
Evidence standardA written definition of "new" and "made" agreed before the first synthesis.
DataDeposit of raw diffraction files, code and failed attempts with the thesis.
SustainabilityEarly life-cycle screen of the shortlist, with its uncertainty stated.

Why sustainability belongs at the start

Most materials are assessed for environmental impact only once composition and synthesis route are fixed, when changes are costly. Woods-Robinson and Trewartha (2026) argue that life-cycle thinking should enter at the lowest technology readiness levels, where design freedom is greatest, and that incomplete data at that stage should be treated as uncertainty to explore rather than a reason to wait. For a doctoral project this means ranking candidates not only by predicted stability and performance but also by the energy their processing needs, the supply of their precursors and what happens at end of life. A candidate that is easy to compute but hard or costly to make sustainably is a weaker contribution than one that is slightly less exotic but practical.

What to do now

  • Pick one materials family and one property before reading the AI-discovery literature in depth.
  • Read the critiques alongside the headline papers (Cheetham & Seshadri, 2024; Leeman et al., 2024).
  • Secure access to synthesis and diffraction facilities early, with a co-supervisor on the experimental side.
  • Plan to deposit raw characterisation data and code so others can check your claims.
Your next step

Could this gap become your thesis?

  1. 1Read the evidence and the open questions
  2. 2Talk to a mentor about fit, data and method
  3. 3Build a defensible proposal
See proposal support

Questions scholars ask

Do I need to build my own AI model?

No. Testing existing models against laboratory outcomes is a valuable contribution, and benchmarks show that evaluation choices matter (Riebesell et al., 2025).

Is the 2021 list of trends still useful?

As background on the areas, yes, but its sources predate the recent shift to large-scale prediction and automated synthesis, so a new proposal should start from 2023 to 2026 work (Merchant et al., 2023; Park et al., 2026).

Can this be done without an autonomous laboratory?

Yes. Conventional synthesis with careful characterisation is enough to test predictions; automation speeds the work but does not replace expert checks (Leeman et al., 2024; Tom et al., 2024).

Sources

  1. Cheetham, A. K., & Seshadri, R. (2024). Artificial intelligence driving materials discovery? Perspective on the article: Scaling deep learning for materials discovery. Chemistry of Materials, 36(8), 3490–3495. https://doi.org/10.1021/acs.chemmater.4c00643
  2. Leeman, J., Liu, Y., Stiles, J., Lee, S. B., Bhatt, P., Schoop, L. M., & Palgrave, R. G. (2024). Challenges in high-throughput inorganic materials prediction and autonomous synthesis. PRX Energy, 3(1), 011002. https://doi.org/10.1103/PRXEnergy.3.011002
  3. Merchant, A., Batzner, S., Schoenholz, S. S., Aykol, M., Cheon, G., & Cubuk, E. D. (2023). Scaling deep learning for materials discovery. Nature, 624(7990), 80–85. https://doi.org/10.1038/s41586-023-06735-9
  4. Park, H., Mastej, K. O., Detrattanawichai, P., Nduma, R., & Walsh, A. (2026). Closing the synthesis gap in computational materials design. Nature Synthesis, 5(5), 651–659. https://doi.org/10.1038/s44160-026-01027-2
  5. Riebesell, J., Goodall, R. E. A., Benner, P., Chiang, Y., Deng, B., Ceder, G., Asta, M., Lee, A. A., Jain, A., & Persson, K. A. (2025). A framework to evaluate machine learning crystal stability predictions. Nature Machine Intelligence, 7, 836–847. https://doi.org/10.1038/s42256-025-01055-1
  6. Szymanski, N. J., Rendy, B., Fei, Y., Kumar, R. E., He, T., Milsted, D., McDermott, M. J., Gallant, M., Cubuk, E. D., Merchant, A., Kim, H., Jain, A., Bartel, C. J., Persson, K., Zeng, Y., & Ceder, G. (2023). An autonomous laboratory for the accelerated synthesis of inorganic materials. Nature, 624(7990), 86–91. https://doi.org/10.1038/s41586-023-06734-w
  7. Szymanski, N. J., Rendy, B., Fei, Y., Kumar, R. E., He, T., Milsted, D., McDermott, M. J., Gallant, M., Cubuk, E. D., Merchant, A., Kim, H., Jain, A., Bartel, C. J., Persson, K., Zeng, Y., & Ceder, G. (2026). Author correction: An autonomous laboratory for the accelerated synthesis of inorganic materials. Nature, 650, E1. https://doi.org/10.1038/s41586-025-09992-y
  8. Tom, G., Schmid, S. P., Baird, S. G., Cao, Y., Darvish, K., Hao, H., Lo, S., Pablo-García, S., Rajaonson, E. M., Skreta, M., Yoshikawa, N., Corapi, S., Akkoc, G. D., Strieth-Kalthoff, F., Seifrid, M., & Aspuru-Guzik, A. (2024). Self-driving laboratories for chemistry and materials science. Chemical Reviews, 124(16), 9633–9732. https://doi.org/10.1021/acs.chemrev.4c00055
  9. Woods-Robinson, R., & Trewartha, A. (2026). Sustainability-informed materials design (arXiv:2604.22987) [Preprint]. arXiv. https://arxiv.org/abs/2604.22987
  10. Zeni, C., Pinsler, R., Zügner, D., Fowler, A., Horton, M., Fu, X., Wang, Z., Shysheya, A., Crabbé, J., Ueda, S., Sordillo, R., Sun, L., Smith, J., Nguyen, B., Schulz, H., Lewis, S., Huang, C.-W., Lu, Z., Zhou, Y., Yang, H., Hao, H., Li, J., Yang, C., Li, W., Tomioka, R., & Xie, T. (2025). A generative model for inorganic materials design. Nature, 639(8055), 624–632. https://doi.org/10.1038/s41586-025-08628-5

How to cite this page

Choose your referencing style, then copy.

PhD Research Lab Insights. (2021, July 19). Materials science future directions: from predicted materials to ones we can actually make. PhD Research Lab. https://www.phdassistance.com/research-methodology/current-state-and-prospects-of-materials-science-research/

PhD Research Lab Insights. “Materials Science Future Directions: From Predicted Materials to Ones We Can Actually Make.” PhD Research Lab, 19 July 2021, www.phdassistance.com/research-methodology/current-state-and-prospects-of-materials-science-research. Accessed [access date].

PhD Research Lab Insights (2021) Materials science future directions: from predicted materials to ones we can actually make. PhD Research Lab. Available at: https://www.phdassistance.com/research-methodology/current-state-and-prospects-of-materials-science-research/ (Accessed: [access date]).

PhD Research Lab Insights. “Materials Science Future Directions: From Predicted Materials to Ones We Can Actually Make.” PhD Research Lab. July 19, 2021. https://www.phdassistance.com/research-methodology/current-state-and-prospects-of-materials-science-research.

PhD Research Lab Insights, ‘Materials Science Future Directions: From Predicted Materials to Ones We Can Actually Make’ (PhD Research Lab, 19 July 2021) <https://www.phdassistance.com/research-methodology/current-state-and-prospects-of-materials-science-research> accessed [access date].

PhD Research Lab Insights. Materials science future directions: from predicted materials to ones we can actually make [Internet]. PhD Research Lab; 2021 Jul 19 [cited [access date]]. Available from: https://www.phdassistance.com/research-methodology/current-state-and-prospects-of-materials-science-research/

PhD Research Lab Insights, “Materials science future directions: from predicted materials to ones we can actually make,” PhD Research Lab, July 19, 2021. [Online]. Available: https://www.phdassistance.com/research-methodology/current-state-and-prospects-of-materials-science-research. [Accessed: [access date]].

@online{current_state_and_prospects_of_materials_science_research_2021,
  author = {{PhD Research Lab Insights}},
  title = {{Materials science future directions: from predicted materials to ones we can actually make}},
  organization = {PhD Research Lab},
  date = {2021-07-19},
  url = {https://www.phdassistance.com/research-methodology/current-state-and-prospects-of-materials-science-research},
  urldate = {[access date]}
}

  • Materials science
  • Materials discovery
  • Machine learning
  • Self-driving laboratories
  • Synthesis
  • Sustainability
Materials science future directions: from predicted materials to ones we can actually makeResearch gap · Engineering & technology
Proposal support