AIによる材料発見——ふるいには効くが発見者にはまだ早い

GNoME（Google DeepMind）, A. Merchant, E. D. Cubuk ら「Scaling deep learning for materials discovery」, Nature（2023年11月）。先行研究（Materials Project）を基準にすれば安定な構造が220万個、うち38万1000件が更新後の凸包の上に残る新規材料——原典の逐語は Of these, 381,000 entries live on the updated convex hull as newly discovered materials で、38万1000は220万の部分集合である。ICSD の実験構造と一致したのは736件で、原典は要旨でこれを 736 have already been independently experimentally realized と書く。ただしこの「独立に」はGNoMEと無関係な他者の実験でという意味であって、予測が実験に先行したという意味ではない——著者らの Methods は 184 of these structures correspond to novel discoveries since the start of the project と書いており、予測が実験に先行したと読めるのはこの184件だ。原典はこの38万1000についても、将来の発見で凸包から落ちうると自ら書いている——Consistent with other literature on structure prediction, the GNoME materials could be bumped off the convex hull by future discoveries, similar to how GNoME displaces at least 5,000 'stable' materials from the Materials Project and the OQMD。新規性の側の主張は leading to more than 45,500 novel prototypes in Fig. 2c (a 5.6 times increase from 8,000 of the Materials Project), which could not have arisen from full substitutions or prototype enumeration（プロトタイプの計数は XtalFinder による）。ふるいの性能は的中率で improve the precision of stable predictions (hit rate) to above 80% with structure and 33% per 100 trials with composition only, compared with 1% in previous work と書かれ、能動学習の出発点は start at less than 6% and 3% だった。この的中率は DFT に対する安定判定の精度であって、合成可能性ではない。 https://www.nature.com/articles/s41586-023-06735-9 ↩ ↩² ↩³ ↩⁴ ↩⁵

GNoME批判：A. K. Cheetham & R. Seshadri「Artificial Intelligence Driving Materials Discovery? Perspective on the Article: Scaling Deep Learning for Materials Discovery」, Chemistry of Materials（ACS, 2024）。「新規性・確からしさ・有用性という三拍子を満たす化合物の証拠は乏しい」（逐語: scant evidence for compounds that fulfill the trifecta of novelty, credibility, and utility）と結論。繰り返し挙げるのは、既知構造と実質同一で対称性だけが低い登録（These structures are virtually identical, aside from the lower symmetry in the GNoME entry／pseudosymmetry that is present in many of the entries）と、many of the entries are based upon the ordering of metal ions that are unlikely to be ordered in the real world の二点。無作為抽出（selecting 10 compounds from among the 384,870 database entries）については We were able to identify the structure of every one of the 10 Stable Structure entries in the ICSD database, albeit usually with a space group of higher symmetry than that in the AI database、「材料」という語については Since no functionality has been demonstrated for the 384,870 compositions in the Stable Structure database, they cannot yet be regarded as materials。空間群の偏りは数でも示されていて、In fact, the top four space groups in the Stable Structure database are all non-centrosymmetric and account for ∼34% of all the structures. By contrast, in the ICSD there is only one non-centrosymmetric space group in the top 24 and it accounts for only 1% of all structures。乱れの見落としが A-Lab 批判と共通することも著者ら自身が書いている（The main reason for the disparity is due to the frequent prediction of structures with atoms ordered on distinct crystallographic sites that—in most cases—are likely to be disordered／We also note that this issue has been highlighted + in a commentary on a different Nature article on robotic/AI-based materials discovery。参照先は Leeman ら PRX Energy 3, 011002 と Szymanski ら Nature 624, 86–91）。放射性元素については In fact, there are 18,138 compounds of such radioactive elements in the large Stable Structure database, including those of Pm, Ac, and Pa ／ There are a further 23,529 entries for compounds containing the highly radioactive elements Tc, Np, and Pu、結論節は These include the elimination of a large number of radioactive materials that are unlikely to have any utility in the materials world と書く。濾し分けの負担については What is now needed is greater effort to connect the predictions to what is already known in the literature to filter out the many candidates that are not truly novel. It is impractical to do this manually with a list of almost 400,000 new compositions。一方で the computational approach delivers credible overall compositions, which gives us confidence that the underlying approach is sound とも書いており、手法そのものを否定してはいない。 https://pubs.acs.org/doi/10.1021/acs.chemmater.4c00643 ↩ ↩²

A-Lab（バークレー）：N. J. Szymanski ら「An autonomous laboratory for the accelerated synthesis of novel materials」, Nature 624, 86–91（2023年11月）。当初「17日・58標的のうち41新化合物」。この表題は原版のもので、2026年の訂正で “novel” が “inorganic” に改められた。このURLが現在返すのは訂正後の版であり、表題は An autonomous laboratory for the accelerated synthesis of inorganic materials、要旨は realized 36 compounds from a set of 57 targets になっている。成功の定義は原版の Methods に Such cases were still considered successful as long as the disordered version of the target retained the same crystal structure and overall composition as the ordered version と在り、訂正後の版はこれを本文の冒頭節にも we consider a synthesis procedure successful when it forms either an ordered or partially disordered version of its target material として置いた（原版の本文には無い文）。 https://www.nature.com/articles/s41586-023-06734-w ↩ ↩²

A-Lab批判と訂正：(批判) J. Leeman, Y. Liu, J. Stiles, S. B. Lee, P. Bhatt, L. M. Schoop, R. G. Palgrave「Challenges in High-Throughput Inorganic Materials Prediction and Autonomous Synthesis」, PRX Energy 3, 011002（2024）。結論は These errors unfortunately lead to the conclusion that no new materials have been discovered in that work（要旨）／we believe that at time of publication, none of the materials produced by A-lab were new（序論）。成功率については we could agree that three materials were correctly synthesized … In this case, the success rate would be 3/58, or 5%、内訳は two thirds of the claimed successful materials in Szymanski et al. are likely to be known compositionally disordered versions of the predicted ordered compounds。A-Lab の新規性の基準については §II で It seems as if the criterion for novelty of a material is its presence in the Materials Project (MP) and its absence from the ICSD. This criterion is open to criticism と述べたうえで we will mostly use this criterion to assess the novelty of the A-lab synthesis products と断り、§VI の勧告でも This was clearly done by the A-lab paper (absence from the ICSD) but some may take issue with this definition と書く。著者らの応答については Even if those known materials were not in the training set and thus were "new to the A-lab," as claimed by the authors in response, this makes the tile and abstract, which claim "novel materials," highly misleading（tile は原典の誤植）。 https://link.aps.org/doi/10.1103/PRXEnergy.3.011002 ／ (訂正) Nature の Author Correction（タイトルから novel を削除、2026年2月号 650:E1、s41586-025-09992-y）：訓練データ混入の Zn₂Cr₃FeO₈ を除外、報告40成功のうち36が正しく4件はXRDのみでは判定不能、「新規」は予測プラットフォームにとっての新規の意。撤回ではなく訂正。この再解析は著者らが手作業で行ったものだが、訂正は This re-analysis was peer-reviewed post-publication と明記し、謝辞でも the reviewers who assessed the re-analyzed PXRD data post-publication に触れている。 https://www.nature.com/articles/s41586-025-09992-y ↩ ↩² ↩³ ↩⁴

MatterGen（Microsoft Research）：「A generative model for inorganic materials design」, Nature（2025年1月）。Materials Project と Alexandria から再計算した607,683の安定構造（20原子以下）で訓練した拡散モデル（607,683 stable structures with up to 20 atoms recomputed from the Materials Project (MP) and Alexandria datasets）。訓練データは実験で確かめられた物質そのものではなく、DFT で再計算された構造である（Alexandria は計算生成のデータベース）。乱れた実在相が、計算の都合で規則構造として入っていることがある。新規性の判定に使う参照データセットはこれとは別で、ICSD を含む Alex-MP-ICSD, comprising 850,384 unique structures のほうである。実証の4件は無作為に選ばれたのではない——原典は候補を (1) uniqueness and novelty; (2) energy above the hull stability from MatterSim and DFT; (3) phonon stability from MatterSim; and (4) whether the material contains oxygen で絞って75件にし、from which we select four for experimental synthesis after expert inspection と書いている。つまりこの4件は、新規性の篩と専門家の目視を通り抜けたものだった。合成が成功したのはそのうち1件で、Synthesis was successful for one of the four candidates。できたものを著者らは the synthesized material is TaCr2O6, a compositionally disordered version of the ordered structure predicted by MatterGen と記述している。体積弾性率は We also experimentally measure the Young's modulus of the sample by nanoindentation and estimate its bulk modulus using the DFT-computed Poisson ratio of 0.30. The estimated bulk modulus is up to 169 GPa after four measurements (158 ± 11 GPa), in which the maximum of the four measurements is our best estimate given that the experimental powder sample is likely non-compact で、規則構造への DFT 予測値は 222 GPa。 https://www.nature.com/articles/s41586-025-08628-5 ↩ ↩²

MatterGen批判：Mikkel Juelsholt（単著）「Continued challenges in high-throughput materials predictions: MatterGen predicts compounds from the training dataset」, Materials Horizons（RSC, 2026, 査読付き）。要旨は MatterGen was used to predict the novel compound TaCr2O6, which was subsequently synthesised in a disordered form as Ta1/3Cr2/3O2. However, … this is not a novel compound but is identical to the previously reported Ta1/2Cr1/2O2, first described in 1971。同一性は upon a 90° rotation around the crystallographic c-axis, the two structures become identical として示され、組成の未報告は While the exact composition of Ta1/3Cr2/3O2 has never been reported, it is just a point in an already established solid-solution series of Ta Cr oxides と自ら認めたうえで退けている。181 GPa の出所（ref.34）は S. Zhang ら, J. Adv. Ceram. 13, 373–387 (2024) で、熱間プレス焼結した CrTaO₄ の測定値である。訓練データ混入は ICSD collection code 9516 is listed in the reference data set used by MatterGen as entry 956681 と同定子つきで裏づけられている。著者の総括は MatterGen cannot distinguish between its training data and the compounds it predicts. Therefore, human inspection of each predicted compound is needed to ensure novelty, which hinders the rapid materials prediction offered by MatterGen、および MatterGen joins other generative AI tools that struggle with key materials science concepts such as crystallography and atomic disorder。参照データセットとは別に訓練データ側も名指ししている（Furthermore, the training data used for MatterGen contains multiple other Ta Cr oxide structures that are rutile structures expanded to P1／Two of the structures are from the Materials project, mp-753467 in Fig. 2A and mp-756340 in Fig. 2B, while the last is from the Alexandria datasets, agm003743920／(listed as TaCrO4 in the training data)）。性質の面の指摘は Ta1/3Cr2/3O2 could be considered novel because of its properties as Ta1/3Cr2/3O2 was predicted to be a superhard material. Zeni et al. measured the bulk modulus to 169 GPa／However, Ta1/2Cr1/2O2 is an already known superhard material with a bulk modulus of 181 GPa and therefore harder than the compound measured by Zeni et al.／However, this argument fails for two reasons／Secondly, the MatterGen training dataset did not include the hardness of Ta1/2Cr1/2O2, only the material itself／MatterGen, therefore, appears to fail to correctly account for the compositional effects of hardness in this particular class of materials。反論機会については著者自身が Zeni et al., who was not contacted before the publication of this paper と断っている。 https://pubs.rsc.org/en/content/articlelanding/2026/mh/d6mh00268d ↩ ↩²