Learning Materials' Reporting History Improves Synthesizability Rankings

Computational screens and generative models now propose inorganic materials faster than laboratories can make them, so a ranking of which candidates are likely to be makeable has practical value. ^[1] One common approach is positive-unlabeled learning: compositions in the Inorganic Crystal Structure Database (ICSD) count as positives, and everything else counts as unlabeled rather than negative, because a missing composition may simply not have been tried. Earlier models of this kind treat the reported materials as an unordered pile, which discards the order in which chemists found them. ^[1]

The finding

Reporting year as a second training target

Park and Walsh add a second job to an existing synthesizability network. A composition goes in as a 90-dimensional fingerprint built from its elements, passes through a shared stack of five fully connected layers, and feeds two outputs. One scores synthesizability. The other learns a pairwise ranking that gives the later-reported of two materials the higher score, with a margin of 0.1 and a weight of 3.0 in the total objective. ^[1] and Methods

The "discovery date" is a proxy: the earliest publication year attached to a composition's ICSD entries, spanning 1905 to 2024. The authors acknowledge it may not match when a material was first made or recognized. ^[1], Methods and Discussion

What the ranking head changed in the learned space

The authors claim the extra task reorganizes what the network learns. When they fit a ridge regression from the learned representation to reporting year on held-out reported compositions, the rank correlation (Spearman) rose from 0.161 for the baseline to 0.505. ^[1], Fig. 2

The representation also shaped scores for unlabeled compositions. The 10% of held-out unlabeled compositions the probe placed latest in time (1,013) had a higher mean synthesizability score than the 10% placed earliest (0.34 against 0.20), more elements per formula (4.20 against 2.84), and far more oxygen (76.8% of compositions against 19.1%). The authors link the later group to phosphate battery cathodes and oxide perovskites studied in recent decades. Their reading is that long-explored regions have fewer undiscovered compositions left, so a missing entry there more likely means a hard synthesis. The paper does not test that reading directly. ^[1], Fig. 3

The evidence

Table 1: the gain over the plain network and over decoy tasks

The authors built two datasets: 104,837 compositions from the Materials Project (51,010 positive, 53,827 unlabeled), and 408,638 once compositions from the Open Quantum Materials Database are added (51,651 positive, 356,987 unlabeled). Each was split 80/20, and the final scores average a 100-model bootstrap ensemble. ^[1], Methods

True negatives are unknown, so precision cannot be measured directly. The authors set the threshold so the top 8.1% of held-out unlabeled compositions score as positive, using a class prior of 0.081 taken from an earlier study of the baseline, and report a calibrated true positive rate. A random classifier scores 0.081 at that setting. ^[1], Methods; ^[2] On the Materials Project dataset the baseline reached 0.470 and the chronology-aware model 0.524. On the larger dataset the figures were 0.831 and 0.873. ^[1], Table 1

Two decoy versions use the same architecture. Trained on shuffled years, the ranking head gave 0.484 and 0.844. Trained to rank by mean atomic mass, it gave 0.466 and 0.838, slightly below the baseline's 0.470 on the smaller dataset. The authors attribute the gain to the real timeline, since neither decoy matched it. They report the gain over the baseline as significant in both datasets (McNemar's test, p below 10^-46) and say it holds across class priors from 0.01 to 0.5. ^[1], Table 1 and Supplementary S2-1

Which known materials the chronology rescued

On the Materials Project test set the baseline wrongly called 5,408 reported compositions unsynthesizable. The chronology-aware model corrected 1,021 of them and newly misclassified 468 the baseline had right, a net gain of 553. The authors compared how often each element class appeared in the rescued compositions and in the unrescued ones, and f-block elements showed the largest positive difference (15.5 percentage points, against 7.1 for metalloids). The authors tie this to bursts of reporting in rare-earth families, such as 52.5% of lanthanide cuprate entries falling in 1987 to 1996. ^[1], Fig. 4

Training on the past, then ranking what came after

The authors also trained on materials reported up to 2015, 2018 or 2020 and ranked later-reported compositions mixed with 10,765 unlabeled ones. They report higher Precision@K than the baseline at every cutoff, with the largest relative gaps at K of 100 for later cutoffs and greater uncertainty there. The text gives one pair of values: at the 2018 cutoff and K of 100, the baseline scored 0.03 (random selection 0.033) and the chronology-aware model 0.07. Only 807, 363 and 249 later-reported positives were available at the three cutoffs. Fig. 5b plots K of 100 and 1,000 at each cutoff with binomial standard-error bars, and the text reports the chronology-aware model ahead of the baseline in every one of those cases. ^[1], Fig. 5

Screening the unreported space

The authors enumerated 168,388 binary and 49,854,596 ternary compositions of 89 elements and scored those absent from the Materials Project. Charge-balanced ones had higher median scores than charge-unbalanced ones: 0.539 against 0.435 for binaries and 0.552 against 0.401 for ternaries. A literature check of selected high scorers found some already reported, such as YbAu2 in 1966 and URu in 1968. A separate check used 101 binary and 1,990 ternary compositions that are in the ICSD but absent from the Materials Project, which gives a reference set independent of the model. The chronology-aware model classified 37.6% of the binaries as synthesizable against 31.7% for the baseline, and 43.6% of the ternaries against 39.9%. ^[1], Table 2 and Fig. 6; Supplementary S4-3 The supplement adds a qualification for the ternaries: the chronology-aware model also labeled a slightly larger share of all ternary compositions missing from the Materials Project as synthesizable (22.4% against 21.1%), which the authors say may account for part of that difference. When they adjusted its threshold to match the baseline's share, it still identified more ICSD-reported ternaries (826 against 795). ^[1], Supplementary S4-3 and Table S2 The paper reports no new synthesis, so these checks show only that high scores line up with reports that already exist.

Future directions

The sharpest open question is whether reporting year measures discovery or research attention. The authors say it may steer scores toward chemistry that has attracted study and away from viable compositions in neglected regions. ^[1] A proposed comparison is to group held-out reported compositions by how many entries their chemical family has, then compare true positive rate against the baseline within each group. A smaller gain in sparse groups would be consistent with attention playing a part, but it would not establish that, because entry counts also track chemistry, feasibility and database coverage. The comparison could narrow the question without separating discovery from attention outright.

The f-block result suggests a related check, also a proposal. The rescued compositions came from families whose reports cluster in time, so a family-by-family comparison could ask whether the gain disappears for families with an even spread of reporting years.

Recent data are thin. The authors attribute a sharp drop in 2020s entries to database coverage, which leaves 249 later-reported positives at the 2020 cutoff. ^[1], Methods A newer ICSD snapshot would let the time-split test be repeated with more positives and tighter error bars.

The model works on composition alone and cannot separate polymorphs. The authors suggest adding the chronology objective to structure-based and language-model predictors. ^[1] Each would need a comparison of a true-chronology model against the same three controls used here: no ranking head, shuffled years and atomic mass.

The study stops short of any synthesis, so the screen leaves a measurement undone. For compositions that score high and appear in neither the Materials Project, the ICSD nor the literature, what fraction can actually be made in a laboratory, and does that fraction exceed the one for compositions the baseline scores equally high? Because ChronoSyn scores compositions and not crystal structures, any such test would count a success as realization of the composition in some form.