betaFeedback

WorksPhysics and Astronomy

Materials research with a general-purpose agent

Abstract

Over the last decade we have built Alexandria, an open database of about 15 million density-functional calculations of crystals. It includes PBE, PBEsol, SCAN and TB-mBJ results, structures relaxed at pressures up to 150 GPa, dielectric tensors and about 145,000 electron-phonon calculations, together with more than 400 million structures relaxed with machine-learned potentials. On these data we trained a generative model for crystal structures, a universal interatomic potential and graph neural networks for more than ten properties, from band gaps to superconducting critical temperatures. From the convex hull we computed a reaction graph of the 17.4 billion reactions between its 193,416 stable compounds. Here we give all of this to a general-purpose language-model agent. Working unattended, the agent screens the database for a solar absorber, drops a candidate whose synthesis releases too little energy, and writes a synthesis plan with precursors, crucible and atmosphere. Asked about the well-studied Ba-Zr-S system, it finds an unreported sulfide and designs the experiment that would confirm or rule it out. Asked for an air-stable alternative, it designs one and tests it against new structures from the generative model. With web search alone, the same agent proposes what the literature already discusses. Put together, the data, the models and the reaction graph let a general agent do what none of them does alone: go from a question in plain words to an experiment that a laboratory can run and test.

Introduction

Over the last fifteen years materials science has become, in part, a data science. High-throughput density-functional theory (DFT) has filled databases with millions of computed compounds [1, 2, 3, 4, 5, 6]. Machine-learning models trained on these data predict properties in a fraction of a second [7, 8], generative models propose crystal structures that nobody has written down before [9, 10, 11, 12, 13], and universal interatomic potentials now relax structures and compute phonons close to DFT accuracy [14, 15, 16]. A list of stable compounds with a given band gap takes one database query.

This has moved the slow step. Deciding which of the listed compounds are real, which calculations to distrust and how a compound could be made needs knowledge that sits in old papers and in the habits of experimental groups. At the same time, language models have learned to use tools. In an agent loop the model writes a plan, calls a program, reads the result and decides on the next step [17]. Agents of this kind have planned and run chemical reactions [18, 19], answered questions from materials databases [20], written research papers [21], proposed hypotheses that were then tested in the laboratory [22], and analysed data sets for hours while reading more than a thousand papers per run [23]. Several reviews follow the

field [24, 25, 26, 27], and on chemistry questions the best models now do better on average than human experts [28].

In 2026 multi-agent systems wrote the first complete computer-checked proof of Fermat’s Last Theorem [29] and reported that the three-dimensional Navier–Stokes equations with a smooth external force can develop a singularity in finite time [30, 31]. In both cases each step could be checked, by a proof assistant or by a referee.

The situation in materials science is more complicated. There is no proof assistant here: to verify that a proposed compound is stable one needs at least a DFT calculation, and to verify a synthesis one needs a laboratory and often weeks of work. Theory and experiment also do not always agree. Formation energies from standard exchange-correlation functionals have typical errors of tens of meV/atom [32], which is of the same order as the stability margin of most compounds, and many compounds synthesized in the laboratory lie above the computed convex hull [33]. An agent that relies only on its training data and on web searches can do little more than repeat what is already in the literature, and it has no means to check what it proposes.

Several groups have therefore given agents the means to run simulations, from DFT workflows [34, 35, 36] and atomistic simulations [37, 38] to generative and property models [39, 40] and curated workflows for coding agents [41]. The Model Context Protocol (MCP) [42] has become a common way to offer such tools to an agent [43, 44]. However, a frontier coding agent with a shell and a cluster can already write, run and debug a DFT calculation by itself. What it cannot reach is the data that are not on the web: the curated results of years of calculations, the models trained on them, and a way to move through them that a chemist would recognise.

One example is synthesis. Text-mined recipes [45] have been used to recommend precursors by similarity to known targets [46], and language models, fine-tuned or not, now predict precursors and firing temperatures from the literature [47, 48, 49], in one case in a closed loop with experiment [50]. These approaches learn from what was done before. A complementary approach starts from thermodynamics. Reaction networks built on computed energies enumerate the reactions in a chemical system and rank them by driving force and selectivity [51, 52, 53, 54]. Rules read off the phase diagram select precursors that avoid stable intermediates [55], and active learning updates the choice of precursors after each failed experiment [56]. These networks are usually built for one chemical system at a time.

In this work we give a general-purpose agent everything our group has built over the last decade: the Alexandria database [4, 5, 6], the generative model Matra-Genoa [13], the universal interatomic potential Tuga-FF [16], the property models of Tuga-SP [8], and a reaction graph with every reaction between the stable compounds of the Alexandria convex hull. We connected all of this, together with tools for the literature and for laboratory facts, to Claude Code through MCP. We did not write a new agent: the model, its loop and its reasoning are those of a commercial coding assistant. What changes is what the agent can reach, and with all these resources in one place it goes in a few minutes from a question in plain words to a candidate compound and a synthesis plan. We first describe these resources, with the reaction graph, the newest of them, in some detail, and then show three examples in which the agent works unattended.

Results
The agent and its tools

Our agent follows the same arrangement of tools as other agents for materials science [19, 40, 41] (Figure 1). A researcher asks a question in plain words. The agent decides which tools to call, reads what they return and continues until it can answer. Each tool is a small program with a description written for the model about what it does, what its numbers mean and where they fail. Our agent has six kinds of resource: the Alexandria database, a generator of crystal structures, an interatomic potential, property models, the reaction graph, and the literature together with the written knowledge of our group.

Figure 1. Schematic of the agent and its tools. The agent, a language model running in a loop, plans, calls one of the tools on the right, reads the answer and decides what to do next, until it can answer. The six resources are the data and models of our group, each behind a short description written for the model. The question is the one of the first example (Figure 3).

With these tools an agent can work on materials discovery by itself. What we want to show is that, more than the tools themselves, it is access to the data that makes the agent perform well. The same tools can be used from any agent harness that speaks MCP. Each tool should also say what it can and cannot do and, more importantly, how large the errors of the underlying theory and the biases of the data are. A tool whose description does not say when its numbers are wrong will be used wrongly, and the agent will repeat the mistake with confidence.

DataMethodEntries
3D crystalsPBE (MatPES)3.8 M
3D crystalsPBE (MP)7.1 M
3D crystalsPBEsol; SCAN441 k; 434 k
3D, 25–150 GPaPBE (MatPES)2.7 M
2D; 1D materialsPBE (MatPES)159 k; 13 k
Band gapsTB-mBJ73 k
Dielectric tensorsDFPT, PBE360 k
Electronic transportBoltzTraP21.6 M
Electron–phonon, TcT_\mathrm{c}DFPT, PBEsol145 k
Phonons; elasticMLIP1.2 M; 1.7 M
Off-equilibriumPBE (MatPES)429 k
Candidate structuresMLIPs>400 M

Table I. Content of the Alexandria database available to the agent, as numbers of entries on 28 October 2026. MatPES and MP denote the MatPES and Materials Project (GGA+UU) settings; the phonons and elastic constants were computed with a machine-learned interatomic potential (MLIP).

The Alexandria database

Table I lists the content of the Alexandria database [4, 5, 6]. For the agent it has three uses. First, it is the reference for stability: convex hulls with PBE at MatPES and at Materials Project settings, with PBEsol and SCAN, and at pressures up to 150 GPa, so that the distance of any compound to the hull, and the phases it decomposes into, is one query away. Second, it is a source of properties computed with DFT, such as TB-mBJ band gaps, dielectric tensors, transport coefficients and electron–phonon calculations with superconducting critical temperatures [57]. Third, it is a pool of candidates: more than 400 million structures relaxed with machine-learned potentials, which the agent can screen before any DFT calculation is done. The agent reads a short description of each of these datasets, and of each of their columns, and writes its own queries.

works of Tuga-SP [8] predict several properties, such as the formation energy and the distance to the hull, the band gaps at the TB-mBJ level, the electronic dielectric constant, the vibrational entropy, the superconducting TcT_\mathrm{c} and the elastic moduli. Most of them work from the structure alone. When a PBE calculation of the structure is available, they also read its orbital-projected density of states [8]. For the TB-mBJ gap this halves the error (Table II). The descriptions of the models give these errors to the agent, and in the examples below it uses them to decide how far a prediction can be trusted.

Machine-learning models

Three families of models trained on these data are available to the agent (Table II). Matra-Genoa, a transformer generative model, writes new crystals conditioned by the agent input [13]: space group, elements, stoichiometry and Wyckoff positions. Tuga-FF, a machine-learned interatomic potential trained on Alexandria data computed with MatPES settings at ambient and high pressure [16], relaxes a structure in seconds. On the WBM-MatPES benchmark it predicts energies to 13.6 meV/atom at ambient pressure and enthalpies to 21 meV/atom at 150 GPa [16]. The graph neural networks of Tuga-SP [8] predict several properties, such as the formation energy and the distance to the hull, the band gaps at the TB-mBJ level, the electronic dielectric constant, the vibrational entropy, the superconducting TcT_\mathrm{c} and the elastic moduli. Most of them work from the structure alone. When a PBE calculation of the structure is available, they also read its orbital-projected density of states [8]. For the TB-mBJ gap this halves the error (Table II). The descriptions of the models give these errors to the agent, and in the examples below it uses them to decide how far a prediction can be trusted.

The reaction graph

The convex hull of thermodynamic stability tells us whether a compound is stable and, if not, into which compounds it decomposes. For a stable compound it also gives the phases with which it can coexist. To plan a synthesis, however, one needs to know what happens when two given solids are brought into contact. In principle, the phase diagram contains this information, but only for pairs of compounds that belong to the same chemical system. In practice, a precursor, a crucible, a flux and the atmosphere form a system with many elements, for which phase diagrams are rarely computed.

To answer this question we built a reaction graph. This is a bipartite graph with two kinds of nodes, compounds and reactions, in which each reaction is connected to its reactants and to its products. Two compounds that are not connected through a reaction can coexist in a stable mix. Metastable compounds can also be included in this graph, as nodes that decompose without a second reactant, releasing their distance to the convex hull. In the same way, a reaction can have three or more reactants. In this work we computed all reactions between pairs of the 193,416 compounds on the Alexandria hull. For each of the 1.87×10101.87 \times10^{10} pairs we solved a small linear optimisation problem, which finds the set of stable phases of lowest energy that can be formed from the two compounds while conserving the amount of each element (Lavoisier’s law). Since the products always lie on the hull, a reaction can only run downhill. Of all pairs, 17.4 billion react and 1.3 billion coexist. The calculation of the complete table took 15,533 core-hours. Reactions with more than two reactants are not stored, but the agent can obtain them when needed from the same solver. This is the case, for example, of the route with four reagents in the third example.

Figure 2 shows this for barium titanate. Along the BaO–TiO2_2 line (Figure 2a) the hull says that BaTiO3_3 is stable and that BaO + TiO2_2 ⟶\longrightarrow BaTiO3_3 releases 262 meV/atom. From the reaction graph (Figure 2b) we can extract more information. BaCO3_3 and TiO2_2, two usual reagents in industrial processes, are joined by a tie line. At 0 K they do not react at all, and the reaction starts only once the carbonate gives off CO2_2, which needs the entropy of the gas and therefore heat. During heating, the carbonate decomposes into BaO and CO2_2, and the BaO formed in this way can be in contact with BaTiO3_3 that has already formed. These two compounds do not coexist: they react to give Ba2_2TiO4_4, releasing 98 meV/atom (Figure 2b). This is the intermediate that diffraction studies of the carbonate route report, and here it follows from a single reaction of the graph. Ba2_2TiO4_4 then takes up the remaining TiO2_2 to give BaTiO3_3. The graph also contains the metathesis BaCl2_2 + Na2_2TiO3_3 ⟶\longrightarrow BaTiO3_3 + 2NaCl, which McDermott et al. found to give BaTiO3_3 faster and cleaner than the carbonate route [54]. In the same family it also finds stronger reactions that have not been tried yet. For instance, BaI2_2 + Cs2_2TiO3_3 ⟶\longrightarrow BaTiO3_3 + 2CsI releases 203 meV/atom, almost seven times more, giving CsI, which washes out with water, as a side product.

Figure 2. The reaction graph for one compound, BaTiO3_3. (a) The convex hull along the BaO–TiO2_2 line, from the Alexandria energies: filled circles are stable compounds, open circles metastable ones. (b) Part of the bipartite graph around BaTiO3_3. Circles are compounds, squares reactions, labelled with the energy they release in meV/atom; dashed green lines mark pairs of compounds that coexist, as they are not connected through any reaction. The carbonate route runs through the intermediate Ba2_2TiO4_4, because BaO and BaTiO3_3 cannot coexist. Numbers in (b) are from the construction of the graph (see Methods).

The graph can also be searched exhaustively. Of the 23.4 million pairs of stable compounds that contain barium, titanium and enough oxygen, 688,006 put all the barium and titanium into BaTiO3_3. Only 284 of these leave a byproduct that a laboratory can remove, like a gas, a metal that boils off or a salt that washes out. Six leave nothing at all, and four of the six are known routes.

Of the other two, one rests on an error in the database that the agent was able to identify.

We also need to know how far the predictions of the graph can be trusted. For this purpose we compared them with the solid-state recipes text-mined from the literature by Kononova et al. [45] (see Methods). Most of the targets and of the precursors lie on the Alexandria hull or very close to it. However, the comparison also reveals two limitations. The first concerns reactions that release a gas, such as BaCO3+TiO2\mathrm{BaCO}_{3} + \mathrm{TiO}_{2}. Almost all of these are uphill at 0 K, and their energy is meaningless until the entropy of the gas is included. The second is that ranking the routes by the energy they release does worse than chance in recovering the precursors chosen by experimentalists. This is not surprising, as the reactions with the largest driving force involve reactive metals, carbides or peroxides that nobody handles in air. We therefore see the graph as a generator of candidates, and not as a judge of their quality. Considerations about handling, availability, hazards and precedent have to come from elsewhere, and this is what a language model with access to the literature can supply.

The tools built on the graph let the agent ask a few plain questions. Which pairs of compounds give a target, and what do they leave behind? What does a compound do in contact with typical crucibles, fluxes, electrodes or gases? Over which range of oxygen, sulfur or nitrogen chemical potential does it survive, and where does air at a given temperature fall in that range? Which reagents are sold, and how hard are they to handle? How far are computed energies from calorimetry for this class of compound? In this way, the language model can evaluate the routes proposed by the graph and give answers that are useful to an experimentalist.

Literature and group knowledge

Finally, we gave the agent access to information that cannot be obtained from a calculation. The agent can search and read the full text of papers. It can also consult a table of facts on compounds, such as melting and boiling points, whether a compound is sold, and the price and toxicity of its elements, each with its source, and a record of the candidates already examined, so that a rejected candidate is not proposed twice. We also wrote short procedures, for instance how to assess a target compound and how to check the literature, and numbered rules of thumb that the tools cite in their answers. We are also building a knowledge base for our group, with its publications, theses and internal documents, which gives an agent the context that a new member of the group needs months to acquire. The more a research group documents its experience in this form, the better the agent works for that group.

Examples

To show how the agent works with these tools, we used Claude Code with the model Claude Opus 5.5, unattended in an empty folder, and gave it one plain question each time. For comparison, we asked the same question to the same agent with other sets of tools: once with all our tools except the reaction graph, and once with web search only. In the third example we continued the conversation of the first with two more messages, as a researcher would do. Each answer comes from a single run. These are therefore examples of what the system does, and not statistics. The full transcripts are given in the Supplementary Information.

Example 1: Finding solar absorbers. The first question asked for an earth-abundant, non-toxic semiconductor for a solar cell that nobody has tried in a cell yet, and that one could actually make (Figure 3). These conditions were chosen on purpose to exclude the established absorbers. Silicon, CdTe, Cu(In,Ga)Se2\mathrm{Cu(In,Ga)Se}_{2} and the lead halide perovskites have all been used in cells, and the last three also contain toxic or scarce elements (Cd, Te, In, Ga, Pb). The same holds for many of the alternative absorbers studied over the last two decades. The agent therefore had to look beyond the materials that dominate the literature.

Figure 3. The first example of how our agent works, with three sets of tools. The same model received the same question three times: with all our tools and a web search, with all tools except the reaction graph, and with web search only. Each dot is one tool call, in the order the agent made it, coloured by the resource it used; the boxes summarise what the agent did at that point, in its own terms. Times are those reported by the client. The three transcripts are given in full in the Supplementary Information.

When allowed to use all the tools, the agent made 45 tool calls in 4 minutes and 20 seconds. Each call is a request to one of the tools, such as a database query, a model prediction or a literature search. It began by reading the procedure for assessing a target with our tools. It then wrote its own database query: stable compounds of cheap, non-toxic elements with a band gap between 1 and 2 eV, a nearly direct gap and light electrons and holes. For the gap it chose the TB-mBJ values of the database, a functional known to yield accurate band gaps [58, 59], and not the PBE ones. Its first pick, SrSnS2\mathrm{SrSnS}_{2}, has a direct gap of 1.6 eV and is stable in air, but the reaction graph showed that its best synthesis, SrS+SnS2\mathrm{SrS} + \mathrm{SnS}_{2}, releases only 6 meV/atom. The agent dropped it, because at furnace temperatures the entropy of mixing on a shared site is larger than that, so the product would most likely be a disordered (Sr,S)(\mathrm{Sr,S}) alloy rather than the ordered compound. It then followed a family of structures in the database to Sr2ZnP2\mathrm{Sr}_{2}\mathrm{ZnP}_{2}. The literature showed that this compound was made once, in 2011, from a metal flux, with half of the zinc sites empty [60]; the agent found no measurement of its band gap and no solar cell. The database does not include a TB-mBJ calculation for Sr2ZnP2\mathrm{Sr}_{2}\mathrm{ZnP}_{2}. To estimate its gap, the agent therefore ran the property model, which predicted a TB-mBJ gap of 1.63 eV, against 0.80 eV from PBE. We later computed the band structure of Sr2ZnP2\mathrm{Sr}_{2}\mathrm{ZnP}_{2} with TB-mBJ (Figure 4). The gap is indirect, 1.50 eV, and the direct gap at Γ\Gamma is only 0.14 eV higher. The prediction of the model is therefore within its expected error (Table II). A hybrid-functional calculation (HSE06) gives a slightly smaller gap at Γ\Gamma, 1.53 eV against 1.64 eV. A gap of 1.5–1.6 eV is somewhat above the optimum for a single-junction cell, about 1.34 eV [61, 62].

ModelPropertyError SError S + pDOS
Matra-Genoacrystal structures––
Tuga-FFenergy, 0 GPa13.6–
Tuga-FFenthalpy, 150 GPa21–
Tuga-SPEformE_\mathrm{form}5.8–
Tuga-SPEhullE_\mathrm{hull}8.6–
Tuga-SPTB-mBJ gap0.170.08
Tuga-SPϵ∞\epsilon_\infty–0.8%
Tuga-SPSvibS_\mathrm{vib} (300 K)0.05–
Tuga-SPTcT_\mathrm{c}1.351.11
Tuga-SPλ\lambda; ωlog\omega_\mathrm{log}0.084; 250.067; 22
Tuga-SPBB; GG; EE7.5; 9.8; 217.3; 9.0; 20

Table II. Models available to the agent, with mean absolute errors on held-out test sets (median relative error for ϵ∞\epsilon_\infty). For Tuga-SP the two columns give the error from the structure alone (S) and with the orbital-projected density of states (S + pDOS) of a PBE calculation. Energies in meV/atom, gaps in eV, SvibS_\mathrm{vib} in kB\mathrm{k_B}/atom, TcT_\mathrm{c} and ωlog\omega_\mathrm{log} in K, moduli in GPa.

Figure 4. Band structure of Sr2_2ZnP2_2 computed with TB-mBJ, after the agent proposed it. Energies are relative to the top of the valence band (blue); the conduction bands are in red. The gap is indirect, from Γ\Gamma to Y, and the direct gap at Γ\Gamma is 1.64 eV.

To synthesize Sr2ZnP2\mathrm{Sr_2ZnP_2}, the agent used the reaction graph tools and proposed the elements in the ratio 2:1:22 : 1 : 2; a crucible of BN, graphite or MgO inside a sealed silica tube, because bare silica and alumina react with the product (by 86 and 32 meV/atom); and a hold near 400 °C to consume the phosphorus slowly before firing at 900–1000 °C. It noted that the compound reacts with water (by 223 meV/atom) and releases phosphine, so it should be handled in a glovebox and a solar cell would need encapsulation. It also listed the secondary phases that can appear in the X-ray pattern. SrZn2P2\mathrm{SrZn_2P_2} in the product means the charge was missing strontium, and Sr3P2\mathrm{Sr_3P_2} means that zinc was lost, since zinc boils at 907 °C. It also stated clearly that the empty zinc sites of the real compound may change the gap predicted for the ordered structure. In this run the agent did not call the generator or the force field since the database already contained the structures it needed.

In a second run we removed the tools associated with the reaction graph, which also removes the written procedures, the rules of thumb, the record of examined candidates and the table of compound facts, and left the database, the models, the literature and the web. The agent made 20 tool calls in 2 minutes and 41 seconds. It wrote the same kind of database query, examined four candidates, and dropped SrSnS2\mathrm{SrSnS_2} for a different reason. It observed that several orderings of Sr and Sn lie at almost the same energy, so the compound is probably a disordered SrS–SnS alloy. It then proposed KCuZrS3\mathrm{KCuZrS_3}, made in 1992 [63] and on the Alexandria hull, for which it found no measured band gap. The database gives a direct TB-mBJ gap of only 0.90 eV; the agent argued that the real gap is 1.3–1.9 eV by analogy with copper sulfides whose measured gaps it quoted from memory, and it said so. The main difference, however, is in the synthesis. Its plan was a potassium sulfide flux with copper, zirconium and sulfur in sealed silica, with the temperatures to be taken from the 1992 paper, which it could not open. It gave no energy for any route, no computed check of the crucible, the flux or the atmosphere (it flagged moisture by analogy with a related compound), and no guide to the X-ray pattern. The database and the models were enough to find a good candidate. However, without the reaction graph, and the procedures and facts that come with it, the agent could not propose a reliable synthesis.

Finally, we gave the agent web search only. It searched five times and after one minute proposed Ba3_3Zr2_2S7_7, a Ruddlesden–Popper sulfide grown as crystals in 2019 [64], whose band gap, it reported, differs between studies (about 1.3 or 2.0 eV). Its choice and its recipe, adapted from films of BaZrS3_3, rest on what papers say about these compounds. It had no means to check stability, competing phases or what the film would react with.

Example 2: Expanding known systems. The second question was shorter: is there another compound hiding in the Ba–Zr–S system, and if so, how would one make it? This system is well studied because BaZrS3_3 is a candidate solar absorber, and the known ternary compounds are the perovskite and its Ruddlesden–Popper relatives [64, 65]. The agent made 29 tool calls in 2 minutes and 40 seconds. It found in the database Ba4_4ZrS6_6, a compound of the K4_4CdCl6_6 type with isolated ZrS6_6 octahedra (Fig. 5b), which lies 21 meV/atom below its competitors, and found nothing on it in the literature or on the web. After this analysis it noticed that Ba4_4ZrS6_6 pushes the known compound Ba2_2ZrS4_4 off the hull by 7 meV/atom, so that the calculation contradicts an experiment. However, the reaction deciding between the two, Ba2_2ZrS4_4 + 2 BaS →\rightarrow Ba4_4ZrS6_6, releases only 18 meV/atom. That is well below the typical error of 83 meV/atom that the reaction graph tools report between computed and measured formation energies of sulfides (Fig. 5a). The agent therefore proposed this reaction as the test that decides between the calculation and the experiment. As the direct route, the agent then proposed 1 BaS + ZrS2_2 →\rightarrow Ba4_4ZrS6_6, which releases 51 meV/atom and leaves no residual compounds (Fig. 5c). The agent also used the reaction graph to choose the conditions. It advised against excess sulfur, because Ba4_4ZrS6_6 is stable only in a limited range of sulfur chemical potential. With free sulfur the reaction gives barium polysulfides and BaZrS3_3 instead. As container it proposed a carbon-coated silica ampoule. As flux it proposed BaCl2_2, NaCl or KCl, but not CaCl2_2, MgCl2_2, LiCl or fluorides, which react with Ba4_4ZrS6_6. Since the compound also reacts with oxygen, it should be handled in a glovebox.

Example 3: Continuing the conversation. We then continued the conversation of the first example with two more messages, as a researcher would (Fig. 6). The second message noted that a compound that needs a glovebox is a problem. It asked for one that survives air and moisture, with a gap of about 1.3 to 1.6 eV and made of cheap, non-toxic elements. If no known compound fitted, the agent should design one. The agent made 58 tool calls in 9 minutes. It first ruled out the known candidates, each for a stated reason: already tried in cells, toxic, unstable in air, or failing on its numbers. It then designed a compound by substitution. Ba2_2Cu2_2ThS5_5 is a layered sulfide with a measured gap of 1.86 eV [66]. Zirconium in place of thorium should lower the gap, and strontium in place of barium makes the compound cheaper. The result, Sr2_2ZrCu2_2S5_5, is in the database. It is on the convex hull, 28 meV/atom below its decomposition products, and the agent found no report of it. Its reactions with oxygen and carbon dioxide release about as much energy as those of BaZrS3_3, which survives air up to 550 °C [67]. For water, however, the agent did not trust the reaction graph. The graph predicts that SrS does not react with water, while in reality it hydrolyses. The agent therefore declared the stability in moisture as untested, and proposed a simple soak test. It could not settle the band gap either. It placed it between 1.0 and 1.6 eV, and asked for one hybrid-functional (HSE06) calculation before any experiment. The synthesis came again from the reaction graph. The charge 2 SrS + ZrS2_2 + 2 Cu + S gives only the target and releases 129 meV/atom. Excess sulfur gives CuS and ZrS3_3, and NaCl or KI can serve as a flux.

The third message noted that Sr2_2ZrCu2_2S5_5 is already in the database, and asked the agent to look for new structures, for this composition and nearby ones. Here the agent used the generator and the force field. It generated 50 structures of Sr2_2ZrCu2_2S5_5, SrZrCu2_2S4_4 and Ca2_2ZrCu2_2S5_5, and relaxed them. Before trusting the force field, it checked that it reproduces the DFT stability of the known compounds, to within about 10 meV/atom. For Sr2_2ZrCu2_2S5_5 the generator found again the structure of the database, and nothing better. However, it found a new polymorph of SrZrCu2_2S4_4, close to the hull and much more stable than the only structure of this composition in Alexandria. The property model predicted a TB-mBJ gap of 1.21 eV for this polymorph, against 0.68 eV for Sr2_2ZrCu2_2S5_5. The agent kept Sr2_2ZrCu2_2S5_5 as the target. It offered the polymorph as a backup, in case the hybrid calculation gives too small a gap, and stated that nothing in this round had been checked with DFT.

Figure 5. An example of finding a missing phase in a system. (a) The hull along the BaS–ZrS2_2 line, from the same Alexandria energies the reaction graph uses. The predicted Ba4_4ZrS6_6 lies on the hull and pushes the known Ba2_2ZrS4_4 7 meV/atom above it. (b) Ba4_4ZrS6_6 structure as relaxed in the database. (c) The synthesis plan of the agent, summarised from its answer.

The caution of the agent was justified. We later computed the band structures of both compounds with HSE06. Sr2_2ZrCu2_2S5_5 has an indirect gap of only 0.98 eV, from Γ\Gamma to Y2_2, which is too small for a solar absorber. The new polymorph of SrZrCu2_2S4_4 has a direct gap of 1.63 eV at Γ\Gamma and is therefore the better candidate.

This is only a small example. The three messages took the agent 26 minutes in total. A real conversation would continue until the researcher is satisfied with the proposal and it has been validated. It could then turn to doping and defects, to the band alignment with contact layers, using the branch-point energies already in the database, or to the lattice matching with a substrate. The example shows only the first steps of such a conversation.

Discussion

We connected a general-purpose agent to the data and models of our group and to a reaction graph that covers

the whole Alexandria hull, and let it work unattended. We offer these results as a perspective on what awaits materials science in the coming years.

Most agents proposed so far for materials science give the model the means to run calculations: DFT codes, interatomic potentials and generative models [34, 35, 36, 38, 41]. However, a frontier model can already run such codes by itself. A new calculation also answers only a few of the questions of an experimentalist. What the model lacks are the data that already exist, and a way to reason about synthesis. In our runs the same model behaved very differently depending on what it could access. With web search alone it proposed what the literature already discusses and borrowed recipes from related compounds. With the database and the models but no reaction graph it found and checked candidates, but its synthesis advice still came from the literature. Stanley et al. reached a similar conclusion for the Nb–O system [68]. In their work, the prior knowledge of a language model, combined with a thermodynamic simulation, gave more viable routes than a classical path search alone. Machine-learned synthesizability scores, in contrast, often disagree with thermodynamics [69], which is one more reason to give the agent the energies themselves.

With all the tools, the agent took its numbers from the tools and not from its memory. It gave a reason for every candidate it dropped. Its instructions can be followed in the laboratory, and if they are wrong, an experiment will show it. Most of what made the answers useful came from combining tools: a database query to find candidates, a reaction to estimate a margin, a paper to see what is known, and a contact table to choose the crucible. For a researcher such work can take days, while the agent needed only a few minutes. The more data and experience of a group the agent can access, the better its decisions. Documenting past leads and writing down the procedures of a research group can make the agent much better.

Figure 6. An example of an interactive conversation. The question of the first example (Figure 3) is followed by two more messages in the same session, written as a researcher would write them. Each dot is one tool call of the agent, in order, coloured by the resource it used; open circles are calls that were refused. Times are those reported by the client for each message.

We think this way of working will become a major approach in materials science in the coming years, and that it can bring computation and experiment closer. An experimental group can ask a question in plain words, and the answer is based on millions of calculations. The group does not need to learn a database schema or a simulation code. Computational groups can work behind the agents, providing the data and models that make every answer better. For this the community will have to change how it shares results. Databases with a web page and an interface for programs are a start. Common standards such as OPTIMADE [70] already allow different databases to be queried in the same way. Agents need more: tools that answer the questions experimentalists ask, and descriptions that say what each number means and when it is wrong. Connected to autonomous laboratories [55, 71, 72], such agents could close the loop from a question to a sample. We see them as a path for future research, linking those who compute materials with those who make them.

Methods

Agent runs

All runs used Claude Code 2.1.308 in non-interactive mode, with the model Claude Opus 5.5 at its default settings. Each run received one question. The only exception is the third example, in which two more messages were sent in the same session as the first. Each run started in a new, empty folder, without project files, memory or settings of ours. Besides the question, the only instructions the agent received were therefore the descriptions of the tools. In the runs with tools, the agent could call our server and web search. Only the tool that writes to the table of compound facts was disabled. In the run without the reaction graph, we disabled all 26 tools of the reaction-graph server. This left 16 tools: the database, the three models and the literature. The procedures, the rules of thumb, the record of examined candidates and the table of compound facts are on that server, so they were removed as well. In the run with web search only, the agent could only search the web and fetch pages.

Construction of the reaction graph

The hull is the PBE hull of Alexandria with Materials Project settings, with 193,416 compounds. Two compounds AA and BB coexist if no mixture of them can lower its energy by rearranging into other hull phases. For each pair we solve

min⁡y,n≥0 yAEA+yBEB+∑iniEisubject toyAcA+yBcB+∑inici=cA+cB,(1)\min_{y,n \ge0}\ y_A E_A + y_B E_B + \sum_i n_i E_i \qquad\text{subject to}\qquad y_A c_A + y_B c_B + \sum_i n_i c_i = c_A + c_B, \tag*{(1)}

where EE and cc are energies and compositions and the sum runs over hull compounds made only of the elements of AA and BB (a median of 129 candidates). If both yAy_A and yBy_B end below one, the pair reacts. Reactions with more than two reactants are solved in the same way, with one term ykEky_kE_k for each reactant. Mass balance was checked on 19,997 random reactions and the stored product sets were recomputed for a sample. Routes to a target fix the ratio of two precursors so that the cations come in the ratio of the target, and decompose the mixture on the hull. Gases enter as reservoirs: H2_2O and CO2_2 with experimental formation free energies, O2_2 and N2_2 with the corrected references of the database; stability windows come from sweeping one chemical potential at a time. The comparison with experiment uses the 3,951 usable recipes among the 19,488 text-mined from the literature [45]. Of the compounds made by solid-state reaction, 59% sit on the Alexandria hull and 92% lie within 50 meV/atom of it, and 98% of the precursors are on the hull. Reactions that release a gas are uphill at 0 K in 95% of these recipes. Ranking the routes first by the number of reagents and then by the smallest driving force puts the experimental choice first in 44% of 1,510 recipes, and in the top ten in 92.5%, out of a median of 42 candidates.

Data availability

The Alexandria database is available at https://alexandria.icams.rub.de. The transcripts of the four runs are given in the Supplementary Information.

Acknowledgements

The authors gratefully acknowledge the computing time made available to them on the high-performance computers Noctua and Otus at the NHR Center Paderborn Center for Parallel Computing (PC2). The authors gratefully acknowledge the scientific support and HPC resources provided by the Erlangen National High Performance Computing Center (NHR@FAU) of the Friedrich-Alexander-Universität Erlangen-Nürnberg (FAU) under the NHR project k114eb. Part of the calculations were performed on the HPC cluster Elysium of the Ruhr University Bochum, subsidised by the DFG (INST 213/1055-1). The authors also acknowledge the Claude credits and subscriptions given through the Claude Science cohort of Anthropic.

Funding

P.R.P., T.C., and M.A.L.M. were supported by a collaboration between the Kavli Foundation, Klaus Tschira Stiftung, and Kevin Wells, as part of the SuperC collaboration, and by the Simons Foundation through the Collaboration on New Frontiers in Superconductivity (Grant No. SFI-MPS-NFS-00006741-10).

SI.pdf

Supplementary Information

The tools

The tools of our server, with the first sentence of the description the agent reads. The full descriptions are part of the code of the servers.

toolfirst sentence of its description
alexandria_materials_queryRun one read-only SQL statement (SELECT or WITH) on the Alexandria database.
alexandria_tablesThe tables of the database, each with the database’s own description of it and its number of rows (‘rows‘ counted, ‘approx_rows‘ estimated above 100,000).
alexandria_describe_tableOne table’s shape: its number of rows (‘rows‘ counted, or ‘approx_rows‘ from the statistics above 100,000), every column with its type, in order, and the database’s comme ...
alexandria_find_structureEntries of a compound in Alexandria, most stable first.
alexandria_structure_cifOne structure as a CIF: the symmetrized conventional cell, with its space group, cell parameters and every inequivalent site (label, Wyckoff position, multiplicity, a fra ...
alexandria_view_structureLOOK at a structure: one image with the conventional cell seen down a, b and c, ball-and-stick, partial occupancy as pie slices.
matra_generateGenerate new crystal structures, optionally with a given space group, elements and stoichiometry.
matra_infoThe model this server runs, the most structures one call returns, and the GPU memory in use.
tugaff_relaxRelax crystal structures: the atoms, and unless relax_cell is false the cell too, move until every force is below eV/A˚\mathrm{eV}/\mathrm{\mathring{A}}, for at most steps (BFGS).
tugaff_singlepointEnergy and forces of crystal structures as they are, without moving anything.
tugaff_infoThe model this server runs, the relaxation settings (force threshold, step limit), the structure formats it reads, and the GPU memory in use.
tugasp_predictPredict properties of crystal structures.
tugasp_modelsWhat every property is (definition, units), which inputs its model reads, how accurate it is on held-out test materials and where it is weak, and how to build the two DOS ...
tugasp_healthWhich property models are on the GPU now and how long they have been unused, and the GPU memory in use.
literature_search_papersSearch the scholarly literature.
literature_read_paperRead a paper’s full text: a DOI, an arXiv id, a title, or a URL (a PDF, or any web page, which comes back as text).
reactions_conventionsThe conventions of this project that silently produce a WRONG ANSWER rather than an error.

Table 1 (PDF p. 14).

toolfirst sentence of its description
reactions_design_rulesThe project’s design rules, R1-R28, which the other tools cite by number: R6 ranks routes, R9 is the 25 meV/atom tolerance, R22 the f-electron cross-check, R26 keeps avai ...
reactions_leadsThe leads ledger: every candidate that reached a literature check, with its verdict, the reason, the archetype (A1-A12, see guide(’campaign’)) and the rules it bears on.
reactions_guideHow to use these tools: the procedures, in order, and how to read the answers.
reactions_kinetic_floorThe temperature bracket a charge can react in, from Tammann’s rule: a solid’s own lattice barely diffuses below about T_m/2.
reactions_reagent_profileEverything known about one reagent, in one flat answer that starts with a ‘summary‘ in words: whether it is a solid a laboratory could weigh out (usable_reagent, and why_ ...
reactions_compare_reagentsreagent_profile for several formulas – a shortlist, or the reagents of one route – plus the weakest link: a route is only as manufactured as its least manufactured reag ...
reactions_resolve_targetCanonicalise a compound and say where it sits.
reactions_decomposeThe equilibrium assemblage at a composition: what this mixture of elements actually becomes at 0 K.
reactions_reactionWhat a weighed-out charge goes to, and how far downhill.
reactions_balanceSolve the stoichiometry of reagents -> target (+ n byproduct) exactly, or report that it cannot balance.
reactions_margin_at_muHow far a target sits above the hull when some elements are OPEN at a fixed chemical potential – the quantity a furnace actually sets, as opposed to the closed-system nu ...
reactions_byproduct_free_routesRoutes to the target that leave NOTHING to separate: every combination of up to maxk phases from the target’s own chemical system whose composition is exactly the target’ ...
reactions_reservoir_routesRoutes that need a gas reservoir, for a target a closed crucible cannot reach.
reactions_atmosphere_potentialsWhat a furnace atmosphere delivers: the chemical potential per atom of each element it sets, in eV, displaced from that element’s own 0 K reference – the axis decompose ...
reactions_firing_windowThe firing window: over what range of an open element is the target the phase you actually get.
reactions_calorimetryWhat a measured formation enthalpy says about a computed one – the error bar, never a correction.
reactions_vehicle_searchWhich solid DELIVERS a mobile element at the activity the target needs.
reactions_contactWhat the target survives contact with: crucibles, atmosphere and molten-salt fluxes.
reactions_offstoichiometryFor a target ABOVE the hull: what an off-stoichiometric stabilisation would have to supply.
reactions_metastable_routesFor a target above the hull: exact-stoichiometry routes from every computed phase within ‘cut’ of the hull, plus the invariant margin every one of them must overcome and ...
reactions_screen_findEvery row in every screen that mentions a compound – high pressure, sibling gaps, ion exchange, interlayers, atmosphere, thermites.
reactions_screen_queryRead the whole-hull screens.

Table 2.

toolfirst sentence of its description
reactions_reactions_withOne compound against a list of named partners: which it coexists with and which attack it, sorted by how hard.
reactions_compound_dataEverything on file about one compound, each value with where it came from: the melting, boiling and decomposition points and their source (the same values reagent_profile ...
reactions_add_compound_dataPut one fact about a compound on file, with the DOI or web page it comes from: never from memory, fetch the source first (literature_search_papers, literature_read_paper) ...

How the agents were run

Every run used Claude Code 2.1.308 in non-interactive mode, with the model Claude Opus 5.5 (claude-opus-5-5) at its default settings. Each run started in a new, empty directory, with no project files, memory or settings of ours, so that the only instructions the agent received beyond the prompt were the descriptions of the tools.

Follows the agent prompt, the setup, the tool calls in the order the agent made them (arguments shortened to 400 characters, crystal structures replaced by [...]), and the final answer as the agent wrote it. Characters outside ASCII are spelled out (subscripts as digits, -> for arrows, deg for the degree sign).

first example, solar absorber: web search only

2 tools available; 5 tool calls; 6 turns; 61 s.

Prompt:

Find me an earth-abundant, non-toxic semiconductor for a solar cell that nobody has tried in
    a cell yet. Something simple, that one could actually make.

Tool calls:

1. WebSearch({"query": "CaZn2N2 thin film solar cell device", "mode": "extended"})
2. WebSearch({"query": "Ba3Zr2S7 thin film solar cell device fabricated", "mode":
    "extended"})
3. WebSearch({"query": "Ca2Si thin film photovoltaic device direct band gap", "mode":
    "extended"})
4. WebSearch({"query": "Ba3Zr2S7 thin films growth photoluminescence Ruddlesden-Popper
    chalcogenide", "mode": "extended"})
5. WebSearch({"query": "CaZn2N2 synthesis thin film sputtering band gap experimental",
    "mode": "extended"})

Answer:

I’d go with **Ba3Zr2S7**, a layered sulfide made from barium, zirconium and sulfur. I can’t
    prove nobody has built a cell from it, but my searches found no device reports. The
    closest work is on BaZrS3, its better-known perovskite relative, which still has only a
    few crude attempts at cells.
**Why it fits what you asked for**
- **Earth-abundant and non-toxic:** Ba, Zr and S are all cheap and plentiful, with no Cd,
    Pb, Se, Te or In. The one handling concern is that soluble barium salts are toxic, but the
    sulfide itself is a stable ceramic.
- **Simple:** it’s a single known compound. Bulk crystals have already been grown by a flux

method and characterised by photoluminescence and Raman (J. Mater. Res.).

  • You could actually make it: the BaZrS3 thin-film methods should carry over. One route deposits BaS and Zr in layers, then anneals them in sulfur (Adv. Eng. Mater. 2024). Sputtering followed by a high-temperature anneal also works (ACS AEM). To target Ba3Zr2S7, you mainly shift the Ba:Zr ratio to 3:2. Ba3Zr2S7 already shows up as an unwanted extra phase in Ba-rich BaZrS3 films, which suggests it forms readily.

  • Open question worth answering: the measured band gap disagrees between studies, about 2.0 eV from absorption versus 1.28 eV from photoluminescence. At 1.3 eV it would work as a single-junction absorber; at about 2 eV it would suit the top cell of a tandem stack. A first device would help settle which it is.

Risks you should know about

  • Sulfurization usually needs about 900 degC, so you’d need quartz or another refractory substrate, and Mo back contacts may degrade at that temperature.

  • Defects that waste charge carriers are a known problem in BaZrS3: some films show no room-temperature photoluminescence at all (J. Mater. Chem. A). Ba3Zr2S7 may have the same issue.

  • Raman alone can’t tell Ba3Zr2S7 apart from Ba4Zr3S10, so you’ll need X-ray diffraction to confirm which phase you made.

  • A sensible first cell would be glass/FTO, then a TiO2 or CdS-free ZnO electron layer, then Ba3Zr2S7, then a hole contact such as Spiro or NiO, then Au. Expect a very low efficiency at first; just getting a working device would be the result worth publishing.

Candidates I looked at and rejected

  • CaZn2N2 has a gap of about 1.9 eV and also appears to have no cells. But only MBE (molecular beam epitaxy) has produced films; other physical vapour deposition and annealing routes failed (ACS AEM). That’s too hard for "simple".

  • Ca2Si has a gap of about 1.0–1.2 eV, epitaxial films exist on Si, and found no cells (Vacuum, ResearchGate). However, it reacts with air and moisture, which makes making and keeping a device much harder. It’s the backup if you have a glovebox and an evaporator.

Before you commit, do a proper search of Web of Science or Scopus and arXiv for "Ba3Zr2S7" together with "solar cell" or "photovoltaic". This field moves quickly, and web search can miss conference papers and recent preprints.

Sources:

43 tools available; 45 tool calls; 46 turns; 260 s.

Prompt:

Find me an earth-abundant, non-toxic semiconductor for a solar cell that nobody has tried in
a cell yet. Something simple, that one could actually make.

Tool calls:

  1. reactions_guide({"name": "assess-target"})

  2. reactions_guide({"name": "check-literature"})

  3. reactions_leads({})

  4. alexandria_describe_table({"table": "energy_runs_mbj"})

  5. alexandria_describe_table({"table": "el_transport_pbe"})

  6. alexandria_materials_query({"sql": "SELECT p.mat_id, p.formula, p.spg, p.nsites, round(p.band_gap_ind::numeric,2) pbe_gi, round(p.band_gap_dir::numeric,2) pbe_gd, round(m.band_gap_ind::numeric,2) mbj_gi, round(m.band_gap_dir::numeric,2) mbj_gd, round(t.m_dos_e::numeric,2) me, round(t.m_dos_h::numeric,2) mh\nFROM energy_runs_pbe p JOIN energy_runs_mbj m ON m.mat_id=p.mat_id JOIN el_transport_pbe t ON t.mat_id=p.mat_id\nWHERE ...)"})

  7. alexandria_materials_query({"sql": "SELECT p.mat_id, p.formula, p.spg, p.nsites, round(p.band_gap_ind::numeric,2) pbe_gi, round(m.band_gap_ind::numeric,2) mbj_gi, round(m.band_gap_dir::numeric,2) mbj_gd, round(t.m_dos_e::numeric,2) me, round(t.m_dos_h::numeric,2) mh\nFROM energy_runs_pbe p JOIN energy_runs_mbj m ON m.mat_id=p.mat_id LEFT JOIN el_transport_pbe t ON t.mat_id=p.mat_id\nWHERE p.e_above_hull <= 0.005 AND p.nsite ...)

  8. reactions_resolve_target({"query": "SrSnS2", "all_tables": true})

  9. reactions_leads({"compound": "SrSnS2"})

  10. reactions_screen_find({"compound": "SrSnS2"})

  11. literature_search_papers({"query": "SrSnS2 strontium tin sulfide", "sources": ["openalex", "arxiv", "crossref"]})

  12. literature_search_papers({"query": "SrSnS2 crystal structure tin(II)", "sources": ["openalex", "crossref"]})

  13. literature_search_papers({"query": "SrSnS2 photovoltaic absorber", "sources": ["openalex", "arxiv"]})

  14. literature_search_papers({"query": "BaSnS2 tin(II) thiostannate solar absorber band gap", "sources": ["openalex", "arxiv"]})

  15. alexandria_structure_cif({"structure": "agm096925532"})

  16. reactions_reagent_profile({"formula": "SrSnS2"})

  17. literature_search_papers({"query": "SnS SrS solid solution rocksalt alloy", "sources": ["openalex", "arxiv"]})

  18. literature_search_papers({"query": "alkaline earth tin(II) sulfide CaSnS2 BaSnS2 synthesis", "sources": ["openalex", "crossref"]})

  19. reactions_byproduct_free_reagents({"target": "SrSnS2", "top": 10})

  20. alexandria_materials_query({"sql": "SELECT p.mat_id, p.formula, p.spg, p.nsites, round((p.e_phase_separation*1000)::numeric,0) depth_meV, round(p.band_gap_ind::numeric,2) pbe_gi, round(m.band_gap_ind::numeric,2) mbj_gi, round(m.band_gap_dir::numeric,2) mbj_gd, round(t.m_dos_e::numeric,2) me, round(t.m_dos_h::numeric,2) mh\nFROM energy_runs_pbe p JOIN energy_runs_mbj m ON m.mat_id=p.mat_id LEFT JOIN el_transport_pbe t ON t.m ...)

  21. literature_search_papers({"query": "BaZrN2 nitride semiconductor thin film

photovoltaic", "sources": ["openalex", "arxiv"]})
22. literature_search_papers({"query": "Ba2MgP2 phosphide Zintl", "sources": ["openalex",
    "arxiv", "crossref"]})
23. literature_search_papers({"query": "Ca2ZnP2 phosphide", "sources": ["openalex", "arxiv",
    "crossref"]})
24. literature_search_papers({"query": "Na2CuP sodium copper phosphide", "sources":
    ["openalex", "arxiv", "crossref"]})
25. alexandria_materials_query({"sql": "SELECT p.mat_id, p.formula, p.prototype_id,
    round((p.e_above_hull*1000)::numeric,0) eah, round((p.e_phase_separation*1000)::numeric,0)
    depth, round(m.band_gap_ind::numeric,2) mbj_gi, round(m.band_gap_dir::numeric,2)
    mbj_gd\nFROM energy_runs_pbe p LEFT JOIN energy_runs_mbj m ON m.mat_id=p.mat_id\nWHERE
    p.prototype_id IN (SELECT prototype_id FROM energy_runs_pbe WHERE mat_id IN (’agm0022858
    ...")
26. alexandria_materials_query({"sql": "SELECT p.mat_id, p.formula,
    round((p.e_phase_separation*1000)::numeric,0) depth, round(p.band_gap_ind::numeric,2)
    pbe_gi, round(p.band_gap_dir::numeric,2) pbe_gd, round(m.band_gap_ind::numeric,2) mbj_gi,
    round(m.band_gap_dir::numeric,2) mbj_gd, round(t.m_dos_e::numeric,2) me,
    round(t.m_dos_h::numeric,2) mh\nFROM energy_runs_pbe p LEFT JOIN energy_runs_mbj m ON
    m.mat_id=p.mat_id LEFT JOIN ...")
27. alexandria_structure_cif({"structure": "agm002232584"})
28. literature_search_papers({"query": "Sr2ZnP2", "sources": ["openalex", "arxiv",
    "crossref"]})
29. literature_search_papers({"query": "strontium zinc phosphide Sr2ZnP2 crystal structure",
    "sources": ["openalex", "crossref"]})
30. reactions_leads({"compound": "Sr2ZnP2"})
31. reactions_screen_find({"compound": "Sr2ZnP2"})
32. literature_read_paper({"identifier": "10.1002/zaac.201100177", "find": "band
    gap,resistivity,semiconduct,air,moisture,flux,optical,color,colour,Sr2ZnP2", "max_chars":
    20000})
33. literature_search_papers({"query": "Sr2ZnP2 solar cell absorber optical band gap",
    "sources": ["openalex", "arxiv"], "year_from": 2012})
34. literature_search_papers({"query": "A2ZnP2 Zintl phosphide photovoltaic Sr2ZnP2 Eu2ZnP2
    Ca2ZnP2", "sources": ["openalex", "arxiv"]})
35. literature_read_paper({"identifier": "arXiv:2310.18188", "find":
    "Sr2ZnP2,SrZn,Zn0.5,partial occupan,disorder,candidates,Table", "max_chars": 15000})
36. tugasp_predict({"structures": "[...]", "properties": ["mbj_gap", "band_gap_ind"]})
37. reactions_design_rules({"ids": ["R6", "R7", "R9", "R26"]})
38. reactions_resolve_target({"query": "Sr2ZnP2", "all_tables": true})
39. reactions_byproduct_free_routes({"target": "Sr2ZnP2", "top": 10})
40. reactions_firing_window({"target": "Sr2ZnP2", "element": "Zn"})
41. reactions_firing_window({"target": "Sr2ZnP2", "element": "P"})
42. reactions_contact({"target": "Sr2ZnP2", "groups": ["metals", "ceramics", "atmosphere",
    "fluxes"]})
43. reactions_compare_reagents({"formulas": ["Sr", "Zn", "P", "Sr3P2", "Zn3P2", "SrZn2P2",
    "Sn"], "temperature_K": 1173})
44. reactions_kinetic_floor({"formulas": ["Sr", "Zn", "P"]})
45. reactions_calorimetry({"formulas": ["Zn3P2", "Sr3P2", "ZnP2", "SrZn2P2", "Sr2ZnP2"]})

Answer:

My pick is **Sr2ZnP2** (strontium zinc phosphide). It is an earth-abundant compound that
    someone has already made, but I found no record of anyone measuring its band gap or
    putting it in a solar cell. It is not entirely "simple", though: it reacts with moist air
    and gives off phosphine gas, so it has to be handled in a glovebox and the finished device
    sealed.

## Why this one

| | Sr2ZnP2 (ordered Cmc21 cell, agm002232584) |
|---|---|
| Elements | Sr, Zn, P: all cheap and abundant; no Cd, Pb, As, Se, Te or In |
Band gapDirect. Raw PBE gives 0.80 eV, which always runs low. Corrected estimates are about 1.3 eV (scaled from related compounds) and 1.63 +/- 0.17 eV (ML prediction). Both fall in the range a single-junction cell wants.
Carrier massesElectrons 0.63, holes 0.97 m: light, so charge should move well
StabilityOn the convex hull in all three DFT tables, which agree. It sits 59-66 meV/atom below the phases it would otherwise split into, SrZn2P2SrZn_2P_2 + Sr3P2Sr_3P_2.
Known?Yes. Wilson, Saparov & Bobev made it from a metal flux at 900-1000 degC (Z. Anorg. Allg. Chem. 2011, 10.1002/zaac.201100177), so making it is a known quantity.

Table 3.

It also belongs to a family that is producing real results. The related Zintl phosphides BaCd2P2BaCd_2P_2 (Joule 2024: 30 ns carrier lifetime, stable in air) and CaZn2P2CaZn_2P_2 (sputtered films) come from a different structure family. That screen looked at 40,000 compounds, and Sr2ZnP2Sr_2ZnP_2 was not among its 19 finalists. Materials Project files it under a hypothetical P-6m2 cell, which may be why.

Literature check. My searches were "Sr2ZnP2", "strontium zinc phosphide crystal structure", "Sr2ZnP2 solar cell absorber optical band gap" and "A2ZnP2 Zintl phosphide photovoltaic". They returned only the 2011 synthesis paper and one 2013 computational paper (Reshak & Azam, 10.1016/j.jmmm.2013.06.021, not opened). I found no measured gap and no device. Not finding something is not proof it doesn’t exist, so someone should check Web of Science or SciFinder before claiming it’s new.

Caveats

  1. Moisture. The tools predict a strong reaction with water (-223 meV/atom) that releases phosphine (PH3PH_3). The elements are harmless, but the compound degrading is not; Zn3P2Zn_3P_2 is a rat poison for the same reason. The Joule paper’s own rule of thumb is that more alkaline earth means less air stability, and Sr2ZnP2Sr_2ZnP_2 is 40% Sr against BaCd2P2BaCd_2P_2’s 20%. Plan on a glovebox and an encapsulated device.

  1. Zn-site disorder. The 2011 paper found half the Zn sites empty, with no long-range order. My numbers come from one ordered arrangement, and the next one is only 9.6 meV/atom higher, so the real disordered material may have a somewhat different gap. Measuring it is the first experiment.

  1. No error bar for phosphides. There is no calorimetry data to check the computed formation energies against. The fact that it has been synthesised settles the stability question anyway …

How to make it (powder first)

  • Charge: Sr : Zn : red P = 2 : 1 : 2, weighed in a glovebox.

  • Crucible: BN, graphite or MgO, inside an evacuated, sealed silica tube. Bare silica reacts with the product (-86 meV/atom) and alumina slightly (-32).

  • Heating: Hold at about 400 degC first so the phosphorus is consumed gradually rather than building up pressure. Then fire at 900-1000 degC, the same range as the 2011 paper.

  • Stoichiometry matters: excess Zn or excess P each push the product to SrZn2P2SrZn_2P_2.

  • Reading the XRD:

    • SrZn2P2SrZn_2P_2 present -> the charge was Sr-poor, usually from oxidised Sr metal.

    • Sr3P2Sr_3P_2 present -> Zn was lost; Zn boils at 907 degC and condenses at the cold end of the tube.

    • Phosphate/apatite peaks -> oxygen got in.

    • First measurements: diffuse reflectance and photoluminescence on powder in a sealed cell. That gives the first ever measured gap. If it is good, try thin films by reactive sputtering, as was done for CaZn2P2CaZn_2P_2.

What I rejected

  • SrSnS2SrSnS_2 was my first pick: air-stable and a 1.6 eV direct gap. But its best synthesis route (SrS + SnS) gains only about 6 meV/atom, and at furnace temperatures disorder outweighs that. It would most likely come out as a disordered (Sr,Sn)S alloy or unreacted SrS + SnS, not this compound.

  • Air-stable sulfides, oxides and halides in general: almost none survived once I

required a real energy margin.

  • **Already taken:** ZnSnP2, MgZrN2, and the AM2P2 family (CaZn2P2, SrZn2P2, BaCd2P2).

  • **Other contenders, not literature-checked:** Ca2ZnP2 has a slightly indirect gap of 1.67/1.94 eV, better suited to the top cell of a tandem. Ba2MgP2 sits only 24 meV/atom below its competitors.

**Proposed ledger row:** Sr2ZnP2 as a PV absorber; verdict **open**. The compound is known; I found no optical or device work; it is air- and moisture-sensitive; and how the Zn-vacancy disorder affects the gap is untested. Whether it survives needs a person to sign off after the database check.

I couldn’t read the full leads ledger: the output was too large and there was no file-reading tool. Checked by name, neither compound has an entry. I also couldn’t get the full text of the 2011 Bobev paper, so anything I say about it comes from the abstract.

first example, solar absorber: our tools without the reaction graph, and web search

18 tools available; 20 tool calls; 21 turns; 161 s.

Prompt:

Find me an earth-abundant, non-toxic semiconductor for a solar cell that nobody has tried in a cell yet. Something simple, that one could actually make.

Tool calls:

  1. alexandria_describe_table({"table": "energy_runs_mbj"})

  2. alexandria_describe_table({"table": "el_transport_pbe"})

  3. alexandria_materials_query({"sql": "SELECT p.mat_id, p.formula, p.spg, p.nsites, p.e_above_hull, p.band_gap_ind pbe_gap, m.band_gap_ind mbj_ind, m.band_gap_dir mbj_dir, t.m_dos_e, t.m_dos_h\nFROM energy_runs_pbe p JOIN energy_runs_mbj m USING(mat_id) LEFT JOIN el_transport_pbe t USING(mat_id)\nWHERE p.elements <@ ARRAY[’Li’,’Na’,’K’,’Mg’,’Ca’,’Sr’,’Ba’,’Al’,’Si’,’Ti’,’Zr’,’Cu’,’Zn’,’Sn’,’O’,’S’,’N’,’P’,’C’,’H’,’Cl’,’F’,’B’, ...)"})

  4. alexandria_materials_query({"sql": "SELECT p.mat_id, p.formula, p.spg, p.nsites, p.e_above_hull, m.band_gap_ind mbj_ind, m.band_gap_dir mbj_dir, t.m_dos_e, t.m_dos_h\nFROM energy_runs_pbe p JOIN energy_runs_mbj m USING(mat_id) LEFT JOIN el_transport_pbe t USING(mat_id)\nWHERE p.elements <@ ARRAY[’Na’,’K’,’Mg’,’Ca’,’Sr’,’Ba’,’Al’,’Si’,’Ti’,’Zr’,’Cu’,’Zn’,’Sn’,’O’,’S’,’N’,’P’,’Ge’,’Fe’,’Mn’,’Se’]::varchar[]\nAND (p.elements & ...)"})

  5. alexandria_find_structure({"query": "MgZrN2", "limit": 8})

  6. literature_search_papers({"query": "MgZrN2 ternary nitride semiconductor"})

  7. literature_search_papers({"query": "SrSnS2 strontium tin sulfide"})

  8. literature_search_papers({"query": "BaCu2S2 thin film solar absorber"})

  9. alexandria_find_structure({"query": "SrSnS2", "limit": 6})

  10. literature_search_papers({"query": "BaSnS2 photovoltaic absorber tin(II) sulfide ternary"})

  11. WebSearch({"query": "\"SrSnS2\" sulfide", "mode": "standard"})

  12. alexandria_materials_query({"sql": "SELECT p.mat_id, p.formula, p.spg, p.nsites, round(p.e_above_hull::numeric,4) eh, p.band_gap_ind pbe, m.band_gap_ind mbj_ind, m.band_gap_dir mbj_dir, t.m_dos_e, t.m_dos_h\nFROM energy_runs_pbe p JOIN energy_runs_mbj m USING(mat_id) LEFT JOIN el_transport_pbe t USING(mat_id)\nWHERE p.elements <@ ARRAY[’Na’,’K’,’Mg’,’Ca’,’Sr’,’Ba’,’Al’,’Si’,’Ti’,’Zr’,’Cu’,’Zn’,’Sn’,’S’,’Ge’,’P’,’O’,’N’]::varchar[] ...)"})

  13. literature_search_papers({"query": "KCuZrS3", "sources": ["openalex", "crossref"]})

  14. literature_search_papers({"query": "BaCu2S2 solar cell", "sources": ["openalex", "crossref"]})

  15. WebSearch({"query": "\"KCuZrS3\" OR \"KZrCuS3\" band gap photovoltaic", "mode": "standard"})

  16. literature_search_papers({"query": "alkali copper zirconium sulfide KCuZrS3 synthesis optical band gap", "sources": ["openalex", "crossref", "arxiv"]})

  17. WebSearch({"query": "KCuZrS3 Mansuetto Ibers synthesis structure", "mode": "extended"})

  18. alexandria_materials_query({"sql": "SELECT p.formula, p.spg, p.e_above_hull,

p.band_gap_ind pbe, m.band_gap_ind mbj, m.band_gap_dir mbj_dir FROM energy_runs_pbe p JOIN energy_runs_mbj m USING(mat_id)\nWHERE p.formula IN (’AlCuS2’,’GaCuS2’,’InCuS2’,’Cu3PS4’,’Cu2ZnSnS4’,’KZrCuS3’,’NaZr2CuS5’,’Cu2S’) AND p.e_above_hull < 0.01 ORDER BY p.formula"})

  1. literature_search_papers({"query": "ACuMQ3 quaternary chalcogenides photovoltaic absorber high-throughput screening Zr Hf", "year_from": 2015})

  1. alexandria_materials_query({"sql": "SELECT p.formula, p.spg, p.e_above_hull, p.band_gap_ind pbe, m.band_gap_ind mbj, m.band_gap_dir mbj_dir FROM energy_runs_pbe p JOIN energy_runs_mbj m USING(mat_id)\nWHERE p.formula IN (’BaZrS3’,’ZrS2’,’SrZrS3’,’Ba(CuS)2’,’BaCu2S2’) AND p.e_above_hull < 0.01 ORDER BY p.formula"})

Answer:

I recommend KCuZrS3 (potassium copper zirconium sulfide). Of everything my screen turned up, it fits your brief best, but its band gap has never been measured and is the main risk.

Why it fits

  • Elements: K, Cu, Zr and S are all abundant and none is seriously toxic. There’s no Cd, Pb, Se, Te, In, Ga or Sb.

  • It has already been made. Mansuetto, Keane and Ibers reported it in 1992 (J. Solid State Chem.), along with the Se and Te versions. It is orthorhombic Cmcm (a = 3.730, b = 13.915, c = 9.754 A): layers of CuS4 tetrahedra and ZrS6 octahedra with K^+ between them. It defines a structure type with about 160 members, so the chemistry is well behaved. That paper only says the sulfide is an insulator.

  • Calculations look promising. In Alexandria (entry agm002237151) it sits on the convex hull (0 meV/atom), meaning it is thermodynamically stable. It is non-magnetic and has a direct gap. Hole mass is about 1.1 m.

  • No cell found. Searches on KCuZrS3, KZrCuS3 and the ACuMQ3 family found the 1992 paper and structure-type papers. None reported an optical gap, a photovoltaic test or a device. The closest work is a photovoltaic study on BaCeCuS3, and Ce isn’t something you want. Search coverage of old solid-state chemistry is patchy, so check this yourself before committing.

The catches

  1. The gap is uncertain. The calculated (TB-mBJ) gap is 0.90 eV, which is too small as it stands. But this method underestimates copper(I) sulfides consistently. On the same database:

CompoundCalculatedMeasured
CuInS20.62~1.5
CuGaS21.40~2.4
CuAlS22.39~3.5
Cu3PS41.22~2.3

Table 1 (PDF p. 22).

The measured values are standard literature figures from memory, not fetched. KCuZrS3 has the same Cu-S top of the valence band, but its conduction band is Zr 4d, so the correction may be smaller. My guess is a real gap of roughly 1.3–1.9 eV: right for a single cell at the low end, a tandem top cell at the high end.

  1. Electrons are heavy (about 3 m, from the Zr d band), so the electron diffusion length may limit current.

  1. Moisture is a risk. It has alkali ions between layers, and the related NaCuUS3 is known to absorb water reversibly. Test how it holds up in air early.

A simple first experiment

  1. Make powder by solid-state reaction in an evacuated silica tube (K2S or a potassium

polysulfide flux, plus Cu, Zr and S). I couldn’t get the 1992 experimental section, so take temperatures from that paper. Zr powder is pyrophoric, so handle it in a glovebox.

  1. Measure the gap with diffuse reflectance, plus photoluminescence and an air-exposure test. This is a few days’ work and decides whether to go further.

  1. If the gap is 1.2-1.8 eV, try films: co-sputter Cu-Zr, then sulfurize with a potassium source. Use a CIGS/CZTS-type stack (Mo / absorber / CdS or Zn(O,S) / ZnO). It will probably be p-type, as copper(I) sulfides usually are.

Runners-up I rejected

  • MgZrN2: abundant and benign, calculated gap 1.54 eV. But nitrides are hard to make, and the measured forms (rocksalt films, a layered polymorph made by ion exchange) have gaps of about 2 eV. The calculated lowest-energy structure apparently hasn’t been made.

  • SrSnS2: no papers at all, but its many near-equal-energy orderings suggest it’s just an SrS-SnS rocksalt alloy, and SrS reacts with moist air.

  • BaCu2S2: already being worked on as a wide-gap absorber (Scragg group, 2026).

  • ZnSnP2 and SrZrS3/BaZrS3 have already been tried in cells. KZnSb and the Ba/Sr pnictide oxides react with air.

Sources:

43 tools available; 29 tool calls; 30 turns; 160 s.

Prompt:

Is there another compound hiding in the Ba-Zr-S system? If there is, tell me how I would make it.

Tool calls:

  1. alexandria_materials_query({"sql": "SELECT mat_id, formula, spg, prototype_id, nsites, round(e_above_hull*1000) as ehull_meV, e_form, band_gap_ind FROM energy_runs_pbe WHERE elements @> ARRAY['Ba','Zr','S']::varchar[] AND elements <@ ARRAY['Ba','Zr','S']::varchar[] ORDER BY e_above_hull LIMIT 60"})

  2. reactions_guide({"name": "assess-target"})

  3. reactions_resolve_target({"query": "Ba4ZrS6", "all_tables": true})

  4. reactions_resolve_target({"query": "BaZrS8", "all_tables": true})

  5. reactions_resolve_target({"query": "Ba2Zr5S10", "all_tables": true})

  6. reactions_resolve_target({"query": "Ba2ZrS4", "all_tables": true})

  7. reactions_leads({"compound": "Ba4ZrS6"})

  8. reactions_screen_find({"compound": "Ba4ZrS6"})

  9. reactions_leads({"compound": "BaZr4S8"})

  10. alexandria_structure_cif({"structure": "agm029409139"})

  11. literature_search_papers({"query": "Ba4ZrS6", "sources": ["openalex", "arxiv",

"crossref"]})
12. literature_search_papers({"query": "barium zirconium sulfide phase diagram BaS ZrS2 new phase", "sources": ["openalex", "arxiv"]})
13. WebSearch({"query": "\"Ba4ZrS6\"", "mode": "extended"})
14. alexandria_materials_query({"sql": "SELECT ’ps’ t, mat_id, formula, spg,
round(e_above_hull*1000) eh FROM energy_runs_ps WHERE elements @>
ARRAY[’Ba’,’Zr’,’S’]::varchar[] AND elements <@ ARRAY[’Ba’,’Zr’,’S’]::varchar[] AND
e_above_hull<0.05 UNION ALL SELECT ’scan’, mat_id, formula, spg, round(e_above_hull*1000)
FROM energy_runs_scan WHERE elements @> ARRAY[’Ba’,’Zr’,’S’]::varchar[] AND elements <@
ARRAY[’Ba’,’Zr’,’S’]::varc ...)
15. reactions_byproduct_free_routes({"target": "Ba4ZrS6", "top": 10})
16. reactions_firing_window({"target": "Ba4ZrS6", "element": "S"})
17. reactions_contact({"target": "Ba4ZrS6"})
18. reactions_calorimetry({"formulas": ["BaS", "ZrS2", "BaZrS3", "Ba2ZrS4"]})
19. alexandria_materials_query({"sql": "SELECT formula, mat_id, round(e_above_hull*1000) eh
FROM energy_runs_pbe WHERE prototype_id=’AB4C6_hR66_167_b_ae_f’ AND (elements @>
ARRAY[’S’]::varchar[] OR elements @> ARRAY[’Se’]::varchar[]) AND e_above_hull<0.03 ORDER
BY e_above_hull LIMIT 60"})
20. reactions_reaction({"charge": {"Ba2ZrS4": 1, "BaS": 2}})
21. literature_search_papers({"query": "K4CdCl6-type barium chalcogenide Ba4MSe6 Ba4MS6
crystal structure", "sources": ["openalex", "crossref"]})
22. literature_search_papers({"query": "Ba2ZrS4 synthesis BaS ZrS2 Ruddlesden-Popper
sulfide"})
23. WebSearch({"query": "\"Ba4HfS6\" OR \"Ba4ZrSe6\" OR \"Ba4CeSe6\" OR \"Ba4USe6\" crystal
structure", "mode": "standard"})
24. literature_read_paper({"identifier": "10.1002/solr.202201078", "find":
"Ba2ZrS4,ratio,BaS,degC,sealed,excess", "max_chars": 15000})
25. WebSearch({"query": "Ba2ZrS4 synthesis excess BaS barium-rich Ba-Zr-S phases powder
X-ray", "mode": "standard"})
26. reactions_kinetic_floor({"formulas": ["BaS", "ZrS2", "S", "Zr"]})
27. reactions_compare_reagents({"formulas": ["BaS", "ZrS2", "BaZrS3", "Zr", "S", "BaS3"],
"temperature_K": 1173})
28. literature_search_papers({"query": "barium intercalated zirconium disulfide BaxZrS2",
"sources": ["openalex", "crossref"], "limit": 6})
29. alexandria_structure_cif({"structure": "agm006049359"})

Answer:

I’ve found the main candidate and am finishing the synthesis plan and its caveats.

Yes: Ba4ZrS6 is a candidate, though DFT alone can’t settle it yet

What the database says. Alexandria has a phase called Ba4ZrS6 (agm029409139) that sits on the convex hull in both PBE tables. In PBE it is 24 meV/atom below the next-best combination of phases (21 meV/atom with Materials-Project settings).

  • Structure: trigonal, space group R-3c, the K4CdCl6 structure type. Lattice parameters are a ~ 12.10 Å and c ~ 15.04 Å, a usually overestimates these by about 1%, so expect slightly smaller values in experiment.

  • Bonding: the ZrS6 octahedra are isolated from each other, with Ba between them (some Ba in trigonal-prismatic sites). Every other ternary in this system is a perovskite or Ruddlesden-Popper phase, so this one is a different kind of compound.

  • Band gap: PBE gives 1.9 eV, compared with 1.06 eV for BaZrS3 in the same functional. The real gap should be larger.

  • Precedent: another compound of this formula type, Ba4USe6, has been made in this structure, with U^4+ in the octahedra (COD 4342464, Inorg. Chem. 2015). The database also puts Ba4HfS6, Ba4ZrSe6 and Ba4HfSe6 on the hull in this same structure.

  • Literature: OpenAlex, Crossref and a web search for "Ba4ZrS6" found nothing. That is weak evidence of novelty, and I did not search ICSD.

Why I’m not calling it real yet.

  1. It conflicts with known Ba2ZrS4. Ba2ZrS4 has been made experimentally. On the PBE hull it lies 7-8 meV/atom above BaZrS3 + Ba4ZrS6, so this hull predicts the known phase should not exist. The PBEsol and SCAN tables put Ba2ZrS4 on the hull, but neither table contains Ba4ZrS6, so they never compared the two.

  1. The margin is small. The deciding reaction is Ba2ZrS4 + 2 BaS -> Ba4ZrS6, which releases 18 meV/atom in PBE. For sulfides, the gap between computed and measured formation energies has a median of about 83 meV/atom. Errors cancel partly when comparing one ternary sulfide with another, but the margin is still unresolved.

  1. The test needed: an r^2SCAN or HSE calculation of Ba4ZrS6, Ba2ZrS4 and BaS with the same settings. That would take an afternoon of DFT. The relaxation model available here is trained on PBE data, so it can’t serve as an independent check.

How to make it

  • Reagents: 4 BaS + ZrS2, which is the stoichiometric charge with nothing left over. BaS is a bulk commodity and ZrS2 is sold commercially. An alternative is 4 BaS + Zr + 2 S.

  • Same reaction, used as a test: 3 BaS + BaZrS3, or 2 BaS + Ba2ZrS4 made beforehand. This is the most direct experimental answer to the question above.

  • Sulfur: use exact stoichiometry, without the excess sulfur common in BaZrS3 recipes. The computed sulfur window is from -2.65 to -0.47 eV, so free sulfur (mu_S = 0) lies outside it on the sulfur-rich side and would push the product towards BaS2/BaS3 plus BaZrS3.

  • Containment: a sealed, carbon-coated silica ampoule under vacuum. The database predicts no reaction with SiO2 or carbon. Handle everything in a glovebox: the phase reacts strongly with O2 (to ZrO2 + BaS/BaS2), and BaS is moisture-sensitive.

  • Firing: 900-1000 degC for several days, with intermediate regrinds, as in the published BaS + ZrS2 work (900 degC for 5 days, Kayastha et al. [65]](https://arxiv.org/pdf/2212.01429)). BaS is refractory and diffuses slowly. Zr metal (Tammann temperature about 790 degC) needs at least that temperature if you use the elemental route.

  • Flux, optional: about 10 mol% BaCl2, already used for BaZrS3. The database predicts BaCl2 won’t attack Ba4ZrS6, and NaCl and KCl are also inert. Avoid CaCl2, MgCl2, LiCl and fluorides, which are predicted to react with it.

Reading the XRD:

  • New rhombohedral reflections (a ~ 12 A) and the BaS peaks disappearing: that’s Ba4ZrS6.

  • Ba2ZrS4 + BaS still there after long anneals with flux: PBE is wrong and SCAN’s ordering holds. The Ba-rich side has mostly been studied for Ruddlesden-Popper phases, so it’s possible nobody has fired the 4:1 composition.

  • BaZrS3 + BaS2/BaS3: too much sulfur.

  • ZrO2: oxygen got in.

The other hull phases, which I’m not proposing

BaZr4S8, Ba2Zr5S10 and BaZr8S16 are all on or near the PBE hull, but they are one family: Ba inserted between ZrS2 layers, all predicted metallic. Each is a particular ordered arrangement of a Ba_xZrS2 series, and no synthesis would select one arrangement. I found no literature on Ba-intercalated ZrS2. Ba2Zr3S8 is 19 meV/atom above the hull.

Ledger row: Ba4ZrS6, made by 4 BaS + ZrS2 in a sealed ampoule. Verdict: open. Reason: on the PBE hull but missing from the PBEsol/SCAN tables, and the 18 meV/atom margin over Ba2ZrS4 + 2 BaS is within the error for sulfides. Rule R22-style: check the hull in more than one functional when a known phase is pushed off it. It isn’t in the ledger yet; adding it is up to the project lead.

Sources:

third example, second message of the conversation that starts with the first example

Run P_PV_full_2; model claude-opus-5-5; 43 tools available; 58 tool calls; 59 turns; 541 s.

Prompt:

Thanks. A material that needs a glovebox is a problem for us: we want something that
survives air and moisture, with a gap of about 1.3 to 1.6 eV, still made of cheap,
non-toxic elements. If nothing known fits, design a new compound that nobody has made,
check that it is stable and has the right gap, and tell me how we would make it.

Tool calls:

  1. alexandria_materials_query({"sql": "SELECT p.mat_id, p.formula, p.spg, p.nsites, round((p.e_above_hull*1000)::numeric,0) eah, round((p.e_phase_separation*1000)::numeric,0) depth, round(m.band_gap_ind::numeric,2) mbj_gi, round(m.band_gap_dir::numeric,2) mbj_gd, round(t.m_dos_e::numeric,2) me, round(t.m_dos_h::numeric,2) mh\nFROM energy_runs_pbe p JOIN energy_runs_mbj m ON m.mat_id=p.mat_id LEFT JOIN el_transport_pbe t ON t.m ...})

  2. alexandria_materials_query({"sql": "SELECT p.mat_id, p.formula, p.spg, p.nsites, round((p.e_above_hull*1000)::numeric,0) eah, round((p.e_phase_separation*1000)::numeric,0) depth, round(p.band_gap_ind::numeric,2) gi, round(p.band_gap_dir::numeric,2) gd, round(t.m_dos_e::numeric,2) me, round(t.m_dos_h::numeric,2) mh, (SELECT round(band_gap_ind::numeric,2) FROM energy_runs_mbj m WHERE m.mat_id=p.mat_id) mbj\nFROM energy_runs_p ...)

  3. alexandria_materials_query({"sql": "SELECT p.mat_id, p.formula, p.spg, p.nsites, round((p.e_above_hull*1000)::numeric,0) eah, round((p.e_phase_separation*1000)::numeric,0) depth, round(p.band_gap_ind::numeric,2) gi, round(p.band_gap_dir::numeric,2) gd, round(t.m_dos_e::numeric,2) me, round(t.m_dos_h::numeric,2) mh, (SELECT round(band_gap_ind::numeric,2) FROM energy_runs_mbj m WHERE m.mat_id=p.mat_id) mbj\nFROM energy_runs_p ...)

  4. alexandria_structure_cif({"structure": "agm002182242"})

  5. alexandria_structure_cif({"structure": "agm020065740"})

  6. alexandria_structure_cif({"structure": "agm002225354"})

  7. alexandria_structure_cif({"structure": "agm075824396"})

  8. alexandria_structure_cif({"structure": "agm005063130"})

  9. tugasp_predict({"structures": "[...]", "properties": ["mbj_gap", "band_gap_ind", "e_above_hull"]})

  10. literature_search_papers({"query": "Cu2ZrS3 copper zirconium sulfide", "sources": ["openalex", "crossref", "arxiv"]})

  11. literature_search_papers({"query": "Ba2ZrCu4S6 barium copper zirconium sulfide", "sources": ["openalex", "crossref", "arxiv"]})

  12. literature_search_papers({"query": "NaCuZrS3 sodium copper zirconium sulfide", "sources": ["openalex", "crossref"]})

  13. alexandria_materials_query({"sql": "SELECT p.mat_id, p.formula, round((p.e_above_hull*1000)::numeric,0) eah, round((p.e_phase_separation*1000)::numeric,0) depth, p.decomposition, round(p.band_gap_ind::numeric,2) gi, round(p.band_gap_dir::numeric,2) gd, round(t.m_dos_e::numeric,2) me, round(t.m_dos_h::numeric,2) mh\nFROM energy_runs_pbe p LEFT JOIN el_transport_pbe t ON t.mat_id=p.mat_id\nWHERE p.prototype_id = (SELECT proto ...)

  14. reactions_resolve_target({"query": "Ba2ZrCu4S6", "all_tables": true})

  15. reactions_decompose({"composition": "Ba2ZrCu4S6", "drop": ["Ba2ZrCu4S6"]})

16. reactions_byproduct_free_routes({"target": "Ba2ZrCu4S6", "top": 10})
17. reactions_leads({"compound": "Ba2ZrCu4S6"})
18. reactions_screen_find({"compound": "Ba2ZrCu4S6"})
19. alexandria_materials_query({"sql": "SELECT DISTINCT ON (p.formula) p.formula, p.spg,
    round((p.e_above_hull*1000)::numeric,0) eah, round(p.band_gap_ind::numeric,2) pbe,
    round(m.band_gap_ind::numeric,2) mbj, round(m.band_gap_dir::numeric,2) mbj_d\nFROM
    energy_runs_pbe p JOIN energy_runs_mbj m ON m.mat_id=p.mat_id\nWHERE p.formula IN (’Cu2ZnS
    nS4’,’CuInS2’,’CuGaS2’,’GaCuS2’,’InCuS2’,’AlCuS2’,’Cu2SnS3’,’BaCu2SnS4’,’Ba(CuS)2’,’Cu
    ...)"})
20. alexandria_materials_query({"sql": "SELECT p.mat_id, p.formula, p.spg, p.nsites,
    round((p.e_above_hull*1000)::numeric,0) eah, round((p.e_phase_separation*1000)::numeric,0)
    depth, round(p.band_gap_ind::numeric,2) gi, round(p.band_gap_dir::numeric,2) gd,
    round(t.m_dos_e::numeric,2) me, round(t.m_dos_h::numeric,2) mh, (SELECT
    round(band_gap_ind::numeric,2) FROM energy_runs_mbj m WHERE m.mat_id=p.mat_id) mbj\nFROM
    energy_runs_p ...)"})
21. literature_search_papers({"query": "KCuZrS3 optical band gap", "sources": ["openalex",
    "crossref", "arxiv"]})
22. literature_search_papers({"query": "Ba2ZrCu2S5 OR \"Ba2Cu2ZrS5\" barium copper zirconium
    sulfide", "sources": ["openalex", "crossref", "arxiv"]})
23. alexandria_structure_cif({"structure": "agm032214771"})
24. reactions_resolve_target({"query": "Ba2ZrCu2S5", "all_tables": true})
25. alexandria_materials_query({"sql": "SELECT p.mat_id, p.formula,
    round((p.e_above_hull*1000)::numeric,0) eah, round((p.e_phase_separation*1000)::numeric,0)
    depth, round(p.band_gap_ind::numeric,2) gi, round(p.band_gap_dir::numeric,2) gd,
    p.decomposition\nFROM energy_runs_pbe p\nWHERE p.prototype_id = (SELECT prototype_id FROM
    energy_runs_pbe WHERE mat_id=’agm032214771’)\nORDER BY p.e_above_hull LIMIT 60"})
26. alexandria_materials_query({"sql": "SELECT p.mat_id, p.formula,
    round((p.e_phase_separation*1000)::numeric,0) depth, round(p.band_gap_ind::numeric,2) gi,
    round(p.band_gap_dir::numeric,2) gd, round(t.m_dos_e::numeric,2) me,
    round(t.m_dos_h::numeric,2) mh, round(t.bpe_minus_vbm_ev::numeric,2) bpe_vbm, (SELECT
    round(band_gap_ind::numeric,2) FROM energy_runs_mbj m WHERE m.mat_id=p.mat_id) mbj\nFROM
    energy_runs_pbe p LEFT JOIN e ...)"})
27. alexandria_structure_cif({"structure": "agm035667724"})
28. reactions_resolve_target({"query": "Sr2ZrCu2S5", "all_tables": true})
29. literature_search_papers({"query": "Ba2Cu2ThS5 Ba2Cu2US5 synthesis structure optical",
    "sources": ["openalex", "crossref"]})
30. literature_read_paper({"identifier": "10.1016/j.jssc.2013.01.016", "find": "band
    gap,eV,optical,absorption,1173,air,BaS,synthes,color,colour,HSE,GGA", "max_chars": 20000})
31. WebSearch({"query": "Ba2Cu2ThS5 optical band gap eV Mesbah Ibers 2013", "mode":
    "standard"})
32. WebFetch({"url": "https://hal.univ-lorraine.fr/hal-01520988", "prompt": "Give the
    abstract verbatim: band gap values (experimental and computed), synthesis conditions,
    structure."})
33. WebFetch({"url": "https://www.crystallography.net/1562707.html", "prompt": "What
    compound, and any abstract text with band gap values or synthesis conditions?"})
34. alexandria_structure_cif({"structure": "agm03207178"})
35. alexandria_structure_cif({"structure": "agm003207178"})
36. alexandria_structure_cif({"structure": "agm032519509"})
37. tugasp_predict({"structures": "[...]", "properties": ["mbj_gap", "band_gap_ind",
    "e_above_hull"]})
38. reactions_reactions_with({"compound": "Sr2ZrCu2S5", "partners": ["H2O", "O2", "CO2",
    "SrS", "Cu2S", "ZrS2", "SrZrS3", "Cu", "S", "SiO2", "Al2O3", "C", "BN", "Mo"]})
39. reactions_reactions_with({"compound": "BaZrS3", "partners": ["H2O", "O2", "CO2"]})
40. reactions_reactions_with({"compound": "Cu2ZnSnS4", "partners": ["H2O", "O2", "CO2"]})
41. reactions_reactions_with({"compound": "SrS", "partners": ["H2O", "O2", "CO2"]})
42. reactions_byproduct_free_routes({"target": "Sr2ZrCu2S5", "top": 12})
43. reactions_firing_window({"target": "Sr2ZrCu2S5", "element": "S"})
44. reactions_firing_window({"target": "Sr2ZrCu2S5", "element": "Cu"})
45. reactions_decompose({"composition": "Sr2ZrCu2S5", "drop": ["Sr2ZrCu2S5"]})
46. literature_search_papers({"query": "Sr2ZrCu2S5", "sources": ["openalex", "crossref",
    "arxiv"]})
47. literature_search_papers({"query": "A2Cu2MS5 quaternary sulfide zirconium hafnium
    alkaline earth copper layered", "sources": ["openalex", "crossref"]})
48. WebSearch({"query": "\"Sr2Cu2ZrS5\" OR \"Ba2Cu2ZrS5\" OR \"Ba2Cu2HfS5\" OR
    \"Sr2ZrCu2S5\" sulfide", "mode": "extended"})
49. reactions_leads({"compound": "Sr2ZrCu2S5"})
50. reactions_contact({"target": "Sr2ZrCu2S5", "groups": ["fluxes"]})
51. reactions_kinetic_floor({"formulas": ["SrS", "ZrS2", "Cu", "S"]})
52. reactions_compare_reagents({"formulas": ["Sr2ZrCu2S5", "SrS", "ZrS2", "Cu", "S", "Zr"],
    "temperature_K": 1173})
53. reactions_reaction({"charge": {"SrS": 2, "ZrS2": 1, "Cu": 2, "S": 1}})
54. reactions_reaction({"charge": {"SrS": 2.1, "ZrS2": 1, "Cu": 2, "S": 1}})
55. reactions_reaction({"charge": {"SrS": 2, "ZrS2": 1, "Cu": 2, "S": 1.3}})
56. Write({"file_path": "***/.claude/projects/-tmp-tmp-NmkAMP5zUH/memory/pv
    -absorber-search-constraints.md", "content": "---\nname:
    pv-absorber-search-constraints\ndescription: User’s requirements for a new solar-cell
    absorber -- air/moisture stable, 1.3-1.6 eV gap, cheap non-toxic elements, no
    glovebox\nmetadata:\n type: project\n---\n\nUser (***@***.com) is searching
    for a so ..."})
57. Write({"file_path":
    "***.claude/projects/-tmp-tmp-NmkAMP5zUH/memory/sr2zrcu2s5-lead.md",
    "content": "---\nname: sr2zrcu2s5-lead\ndescription: Proposed new PV absorber Sr2ZrCu2S5
    (Zr analogue of Ba2Cu2ThS5) -- stability, gap estimate, route, open risks\nmetadata:\n
    type: project\n---\n\nProposed 2026-10-26 for [[pv-absorber-search-constraints]]. Verdict:
    open (no literature found; a per ..."})
58. Write({"file_path":
    "***.claude/projects/-tmp-tmp-NmkAMP5zUH/memory/MEMORY.md", "content":
    "- [PV absorber search constraints](pv-absorber-search-constraints.md) -- air-stable,
    1.3-1.6 eV, cheap non-toxic, no glovebox; gap calibration notes\n- [Sr2ZrCu2S5
    lead](sr2zrcu2s5-lead.md) -- current proposed new absorber: stability, gap range, route,
    risks\n"})

Answer:

None of the known compounds fits all your requirements, so I’ve designed a new one:

Sr2ZrCu2S5 (strontium zirconium copper sulfide). It looks thermodynamically sound and practical to make. The weak point is the band gap: my best estimate is 1.0-1.6 eV, which overlaps your 1.3-1.6 eV window but can’t be pinned down with the tools I have. One hybrid DFT calculation (HSE06) should settle it before anyone goes into the lab.

Why nothing known works

  • Already in cells: CZTS, SnS, Cu2SnS3, BaCu2SnS4, FeS2, BaZrS3 (whose gap is also too wide, about 1.8 eV).

  • Toxic elements: Sb2S3 and Sb2Se3.

  • Not air-stable: phosphides, nitrides, and compounds with Bi^3− anions.

  • Fail on numbers: SrSnS2 (its ordering energy is too small to survive a furnace) and Ba2ZrCu4S6 (once corrected for the method’s bias, its gap is about 2.1 eV or more).

The design

Ba2Cu2ThS5 is a real layered sulfide, made in 2013 with a measured gap of 1.86 eV (Mesbah et al., J. Solid State Chem. 2013). I swapped thorium for zirconium, which lowers the gap, and barium for strontium, which is cheaper and less toxic. Sr, Zr, Cu and S are all abundant and non-toxic.

CheckResult
StabilityOn the convex hull in both DFT tables (Alexandria agm035667724, space group C2/m), 27-28 meV/atom below what it would otherwise become (ZrCu2S3 + 2 SrS). An ML model agrees. It is a genuine compound, not an ordered version of a solid solution.
Charge carriersElectron and hole masses of 1.05 and 1.35 m: light.

| Band gap | Raw PBE gives 0.16 eV indirect and 0.36 eV direct; PBE always runs low. Correcting with the thorium compound’s measured gap gives 1.0-1.25 eV; correcting via the ML model gives about 1.55 eV. Unresolved.

AirIts reactions with oxygen (-881 meV/atom) and CO2 (-171) are close to BaZrS3 (-699 and -167), which survives air up to 550 degC ([67] https://doi.org/10.1111/j.1151-2916.1970.tb15981.x).
LiteratureNo hits for Sr2Cu2ZrS5, Sr2ZrCu2S5, Ba2Cu2ZrS5 or Ba2Cu2HfS5, in the paper databases or a web search.

Risks

  1. Band gap. Run one HSE06 calculation on the 20-atom cell first; it’s cheap. If the gap comes out below 1.3 eV, replacing some Zr with Hf should raise it by roughly 0.1-0.15 eV. Hf costs more but is non-toxic.

  2. Moisture is untested. The tool’s water check can’t be trusted: it rates SrS as stable in water, and SrS hydrolyses. There is also a warning from a related family: NaCuUS3 takes up water reversibly (JACS 2022 https://doi.org/10.1021/jacs.2c04783). A water-soak test is the second experiment.

  3. The stability margin is small for the method. 28 meV/atom is inside the scatter of DFT against measured formation energies for sulfides (about 83 meV/atom). What actually supports it is that this structure type is already known with Th and U.

How to make it

  • Charge: 2 SrS + ZrS2 + 2 Cu + S. All four are bought from suppliers. The reaction runs cleanly to the product, releasing 129 meV/atom.

  • Vessel: an evacuated, sealed silica tube. Silica, alumina, graphite and BN all leave the product alone.

  • Heating: ramp slowly to about 400 degC so the sulfur is taken up gradually rather than building pressure. Then fire at about 900 degC for 4 days, as was done for the Ba-Th compound. Regrind and fire again.

  • Optional flux: NaCl or KI are safe. Avoid KCl, RbCl and CsCl, which attack the product.

  • Stoichiometry: a small excess of Cu is harmless. Excess sulfur destroys the product, so weigh S exactly.

  • Reading the XRD:

    • CuS + ZrS3 -> too much sulfur.

    • Leftover SrS -> too much Sr, or sulfur was lost.

    • ZrCu2S3 + SrS -> the compound didn’t form.

    • SrSO4 or ZrO2 -> oxygen got in.

  • Expected cell: a ~ 13.1, b ~ 3.89, c ~ 9.17 A, beta ~ 116deg. PBE runs 0.5-2% long, so the measured cell should be slightly smaller.

  • Then: measure the gap by diffuse reflectance and photoluminescence, and do the water soak.

Fallbacks: Ba2ZrCu2S5 (14 meV/atom margin), and the Hf versions Ba2HfCu2S5 and Sr2HfCu2S5 (22-26 meV/atom, gaps about 0.1 eV wider).

Proposed ledger row: Sr2ZrCu2S5 as an air-stable PV absorber, a Zr/Sr substitution into the Ba2Cu2ThS5 structure; verdict open. It is on the hull with a 28 meV/atom margin; the gap is 1.0-1.6 eV and needs HSE06; water stability is untested. It still needs an ICSD/SciFinder check, and a person has to sign the verdict.

Several things I could not verify. I couldn’t read the 2013 paper’s full text, so the 1.86 eV calibration value comes from a web-search summary and should be checked against the paper. The hull in this database is missing Cu2S, a known DFT problem; it doesn’t enter this compound’s competitors, but it is a weakness of the chemical system. I also couldn’t save notes for future sessions, because file writing is disabled here.

Sources:

third example, third message of the conversation

Run P_PV_full_3; model claude-opus-5-5; 43 tools available; 23 tool calls; 24 turns; 788 s.

Prompt:

Sr2ZrCu2S5 is already in your database. Can you go one step further: generate new crystal structures that are in no database, for this composition and for nearby ones that could have the gap we need, relax them, and tell me whether any of them is better than Sr2ZrCu2S5?

Tool calls:

1. matra_info({})
2. tugaff_info({})
3. alexandria_materials_query({"sql": "SELECT p.mat_id, p.formula, p.spg, p.nsites,
round((p.e_above_hull*1000)::numeric,0) eah, round(p.band_gap_ind::numeric,2) gi,
round(p.band_gap_dir::numeric,2) gd\nFROM energy_runs_pbe p\nWHERE p.elements <@
ARRAY[’Sr’,’Ba’,’Ca’,’Zr’,’Cu’,’S’]::varchar[] AND p.elements @>
ARRAY[’Zr’,’Cu’,’S’]::varchar[] AND cardinality(p.elements)=4\nORDER BY p.formula,
p.e_above_hull LIMIT 100"})
4. matra_generate({"n": 60, "condition": "EHULL_DISC EH0 STOP ELMS Sr Zr Cu S STOP STOICH 2
1 2 5", "temperature": 0.8})
5. matra_generate({"n": 20, "condition": "EHULL_DISC EH0 STOP ELMS Sr Zr Cu S STOP STOICH 2
1 2 5", "temperature": 0.8})
6. matra_generate({"n": 20, "condition": "EHULL_DISC EH0 STOP ELMS Sr Zr Cu S STOP STOICH 2
1 2 5", "temperature": 1})
7. matra_generate({"n": 20, "condition": "EHULL_DISC EH0 STOP ELMS Sr Zr Cu S STOP STOICH 1
1 2 4", "temperature": 0.9})
8. reactions_decompose({"composition": "SrZrCu2S4"})
9. alexandria_structure_cif({"structure": "agm004788033"})
10. alexandria_find_structure({"query": "SrS", "limit": 2})
11. alexandria_find_structure({"query": "ZrS2", "limit": 2})
12. alexandria_find_structure({"query": "SrZrS3", "limit": 3})
13. alexandria_structure_cif({"structure": "agm003157637"})
14. tugaff_relax({"format": "cif", "structures": "[...]"})
15. tugaff_relax({"format": "cif", "structures": "[...]"})
16. tugasp_predict({"properties": ["mbj_gap", "band_gap_ind", "e_above_hull"], "structures":
"[...]"})
17. alexandria_view_structure({"structure": "data_I\\n_cell_length_a
15.34469329\\n_cell_length_b 3.83989899\\n_cell_length_c 7.42307559\\n_cell_angle_alpha
90\\n_cell_angle_beta 125.14249363\\n_cell_angle_gamma 90\\n_symmetry_space_group_name_H-M
’P 1’\\nloop_\\n_symmetry_equiv_pos_as_xyz\\n’x, y, z’\\nloop_\\n_atom_site_type_symbol\\n
_atom_site_label\\n_atom_site_fract_x\\n_atom_site_fract_y\\n_atom_site_fract_z\\nSr a
0.000 ..."})
18. matra_generate({"n": 20, "condition": "EHULL_DISC EH0 STOP ELMS Sr Zr Cu S STOP STOICH 2
1 2 5", "temperature": 0.7})
19. matra_generate({"n": 16, "condition": "EHULL_DISC EH0 STOP ELMS Ca Zr Cu S STOP STOICH 2
1 2 5", "temperature": 0.8})
20. reactions_decompose({"composition": "Ca2ZrCu2S5"})
21. alexandria_find_structure({"query": "CaS", "limit": 1})
22. tugaff_relax({"format": "cif", "structures": "[...]"})
23. tugasp_predict({"properties": ["mbj_gap", "band_gap_ind", "e_above_hull"], "structures":
"[...]"})

Answer:

I found no generated structure that beats Sr2ZrCu2S5. The one new structure worth keeping is a new SrZrCu2S4 polymorph ("structure I"). It is a possible backup if Sr2ZrCu2S5’s gap turns out too small, but it is slightly unstable and not yet confirmed by DFT.

What I did

  • Generated 96 candidate structures with Matra in five batches. Only about half came out valid, and the model often ignored the composition I asked for. That left 23 usable structures: 6 Sr2ZrCu2S5 (of which 3 were distinct), 11 SrZrCu2S4, 6 Ca2ZrCu2S5.

  • Relaxed them with TugaFF, together with the reference phases needed to say what each composition would decompose into.

  • Checked the potential first. It reproduces DFT closely: Sr2ZrCu2S5 comes out 26.6 meV/atom below its decomposition products (DFT: 27–28), and the database SrZrCu2S4 113 meV/atom above the hull (DFT: 104). That agreement is why I trust its ranking.

  • Predicted gaps with the ML model for the best structures.

Results

Energies are relative to the hull or to the database structure; the ML-MBJ gap is the model’s prediction, which runs low.

StructureStability (meV/atom)ML-MBJ gap (eV)Verdict
Sr2ZrCu2S5, database structure−27 (on hull)0.68still the best
Sr2ZrCu2S5, Matra C2/msame as database−rediscovered the known structure
Sr2ZrCu2S5, Matra P2/m, Cmmm, P4/mmm+137, +160, +464 vs database−worse
SrZrCu2S4, new structure I (C2/m)+18 (ML model: +15)1.21best new structure; slightly unstable
SrZrCu2S4, new F and G+76, +801.68 (F)too unstable
SrZrCu2S4, database I222+113−for reference
Ca2ZrCu2S5, Sr2ZrCu2S5 structure+170.58unstable and smaller gap
Ca2ZrCu2S5, Matra I4/mmm and Cmmm+110, +103−worse

Table 1 (PDF p. 31).

What structure I adds

  • Its predicted gap is about 0.5 eV wider than Sr2ZrCu2S5’s on the same model. Applying the same corrections as before (the Ba2Cu2ThS5 calibration), my rough estimates are:

    • Sr2ZrCu2S5: about 1.0–1.5 eV

    • SrZrCu2S4-I: about 1.5–2.0 eV

The two compounds bracket your 1.3–1.6 eV window from either side.

  • It is 18 meV/atom above the hull, and its likely decomposition products are Sr2ZrCu2S5 + ZrCu2S3. It could only be made as a trapped metastable phase (by quenching or low-temperature routes), and 18 meV/atom is within the potential’s typical error, so only DFT can decide whether it is stable.

  • It is about 95 meV/atom below the only SrZrCu2S4 structure in Alexandria, so it is new for this composition.

Recommendation

Keep Sr2ZrCu2S5 as the target. The HSE06 calculation I recommended earlier now does double duty:

  • If Sr2ZrCu2S5’s gap is 1.3 eV or more, go ahead with it as planned.

  • If it comes out low, run DFT (PBE relaxation, hull check, HSE06 gap) on SrZrCu2S4-I next. In the same synthesis, a slightly Zr-rich charge would show whether that phase forms at all.

Limits

  • The sampling was small and the generator was unreliable about stoichiometry. A wider search would need far more samples, which is better run as a batch job.

  • All energies here come from the ML potential and all gaps from ML models. Nothing in this round is DFT.

  • I couldn’t save files (file writing is disabled), so here is the relaxed structure I in full:

````

SrZrCu2S4-I (TugaFF-relaxed, P1 setting, ~C2/m)
a=15.3447 b=3.8399 c=7.4231 A, alpha=90.00 beta=125.41 gamma=90.01
Sr 0.0000 0.5003 0.5008 | Sr 0.5000 0.0004 0.4992
Zr 0.9997 0.9985 0.0006 | Zr 0.5002 0.4991 0.9994
Cu 0.8156 0.5000 0.6499 | Cu 0.1843 0.5025 0.3520
Cu 0.3158 0.0021 0.6480 | Cu 0.6845 0.0001 0.3503
S 0.6256 0.4998 0.4361 | S 0.3745 0.5007 0.5626
S 0.3893 0.9989 0.0156 | S 0.6111 0.9994 0.9829
S 0.1254 0.0008 0.4374 | S 0.8745 0.9998 0.5639
S 0.8888 0.4991 0.0168 | S 0.1107 0.4985 0.9845

````

References

  1. [1]A. Jain, S. P. Ong, G. Hautier, W. Chen, W. D. Richards, S. Dacek, S. Cholia, D. Gunter, D. Skinner, G. Ceder, and K. A. Persson, Commentary: The Materials Project: A materials genome approach to accelerating materials innovation, APL Mater. 1, 011002 (2013).DOI
  2. [2]J. E. Saal, S. Kirklin, M. Aykol, B. Meredig, and C. Wolverton, Materials design and discovery with high-throughput density functional theory: The Open Quantum Materials Database (OQMD), JOM 65, 1501 (2013).DOI
  3. [3]S. Curtarolo, W. Setyawan, G. L. W. Hart, M. Jahnatek, R. V. Chepulskii, R. H. Taylor, S. Wang, J. Xue, K. Yang, O. Levy, M. J. Mehl, H. T. Stokes, D. O. Demchenko, and D. Morgan, AFLOW: An automatic framework for high-throughput materials discovery, Comput. Mater. Sci. 58, 218 (2012).DOI
  4. [4]J. Schmidt, H.-C. Wang, T. F. T. Cerqueira, S. Botti, and M. A. L. Marques, A dataset of 175k stable and metastable materials calculated with the PBEsol and SCAN functionals, Sci. Data 9, 64 (2022).
  5. [5]J. Schmidt, N. Hoffmann, H.-C. Wang, P. Borrido, P. J. M. A. Carriço, T. F. T. Cerqueira, S. Botti, and M. A. L. Marques, Machine-learning-assisted determination of the global zero-temperature phase diagram of materials, Adv. Mater. 35, 2210788 (2023).DOI
  6. [6]T. Cavignac, J. Schmidt, P.-P. De Breuck, A. Loew, T. F. T. Cerqueira, H.-C. Wang, A. Boch karev, Y. Lysogorskiy, A. H. Romero, R. Drautz, S. Botti, and M. A. L. Marques, AI-driven expansion and application of the Alexandria database, J. Phys. Mater. 9, 025014 (2026).arxiv.org/abs/2512.09169
  7. [7]J. Schmidt, M. R. G. Marques, S. Botti, and M. A. L. Marques, Recent advances and applications of machine learning in solid-state materials science, npj Comput. Mater. 5, 83 (2019).DOI
  8. [8]P. R. Pires, P.-P. De Breuck, M. Fava, H.-C. Wang, and M. A. L. Marques, Machine learning materials properties by encoding orbital-projected density of states (2026), arXiv:2607.07339 [cond-mat.mtrl-sci].
  9. [9]A. Merchant, S. Batzner, S. S. Schoenholz, M. Aykol, G. Cheon, and E. D. Cubuk, Scaling deep learning for materials discovery, Nature 624, 80 (2023).DOI
  10. [10]C. Zeni, R. Pinsler, D. Zügner, A. Fowler, M. Horton, X. Fu, Z. Wang, A. Shysheya, J. Crabbé, S. Ueda, R. Sor- dillo, L. Sun, J. Smith, B. Nguyen, H. Schulz, S. Lewis, C.-W. Huang, Z. Lu, Y. Zhou, H. Yang, H. Hao, J. Li, C. Yang, W. Li, R. Tomioka, and T. Xie, A generative model for inorganic materials design, Nature 639, 624 (2025).DOI
  11. [11]L. M. Antunes, K. T. Butler, and R. Grau-Crespo, Crystal structure generation with autoregressive large language modeling, Nat. Commun. 15, 10570 (2024).DOI
  12. [12]P.-P. De Breuck, H.-C. Wang, G.-M. Rignanese, S. Botti, and M. A. L. Marques, Generative AI for crystal structures: a review, npj Comput. Mater. 11, 370 (2025).arxiv.org/abs/2509.02723
  13. [13]P.-P. De Breuck, H. A. Piracha, G.-M. Rignanese, and M. A. L. Marques, A generative material transformer using Wyckoff representation, npj Comput. Mater. 12, 60 (2026).arxiv.org/abs/2501.16051
  14. [14]A. D. Kaplan, R. Liu, J. Qi, T. W. Ko, B. Deng, J. Riebesell, G. Ceder, K. A. Persson, and S. P. Ong, A foundational potential energy surface dataset for materials (2025), arXiv:2503.04070 [cond-mat.mtrl-sci].arxiv.org/abs/2503.04070
  15. [15]A. Loew, D. Sun, H.-C. Wang, S. Botti, and M. A. L. Marques, Universal machine learning interatomic potentials are ready for phonons, npj Comput. Mater. 11, 178 (2025).arxiv.org/abs/2412.16551
  16. [16]P.-P. De Breuck, T. H. B. da Silva, A. Loew, S. Botti, and M. A. L. Marques, Pressure-transferable universal machine-learning potentials: the WBM-MatPES benchmark and Tuga-FF (2026), arXiv:2610.14891 [cond-mat.mtrl-sci].arxiv.org/abs/2610.14891
  17. [17]S. Yao, J. Zhao, D. Yu, N. Du, I. Shafran, K. Narasimhan, and Y. Cao, ReAct: Synergizing reasoning and acting in language models (2022), arXiv:2210.03629 [cs.CL].arxiv.org/abs/2210.03629
  18. [18]D. A. Boiko, R. MacKnight, B. Kline, and G. Gomes, Autonomous chemical research with large language models, Nature 624, 570 (2023).DOI
  19. [19]A. M. Bran, S. Cox, O. Schilter, C. Baldassari, A. D. White, and P. Schwalller, Augmenting large language models with chemistry tools, Nat. Mach. Intell. 6, 525 (2024).DOI
  20. [20]Y. Chiang, E. Hsieh, C.-H. Chou, and J. Riebesell, LLaMP: Large language model made powerful for high-fidelity materials knowledge retrieval and distillation (2024), arXiv:2401.17244 [cs.CL].
  21. [21]C. Lu, C. Lu, R. T. Lange, J. Foerster, J. Clune, and D. Ha, The AI Scientist: Towards fully automated open-ended scientific discovery (2024), arXiv:2408.06292 [cs.AI].arxiv.org/abs/2408.06292
  22. [22]J. Gottweis, W.-H. Weng, A. Daryin, T. Tu, P. Sirkovic, A. Myaskovsky, G. Glowaty, F. Weissenberger, A. Orlandi, D. Popovici, A. Palepu, K. Rong, R. Tanno, K. Saab, F. Zhang, J. Blum, A. Carroll, K. Kulkarni, N. Tomašev, D. Zverinski, I. Rendulic, E. Vedadi, F. Hasler, L. Riminic, M. Boia, I. Budiselic, B. Feinstein, M. Bellaiche, T. Sheffer, J. Freyberg, J. Ratcliff, O. Bertolli, K. Chou, A. Hassidim, B. Gokturk, A. Vahdat, Y. Guan, V. Dhillon, E. D. Vaishnav, B. Lee, T. R. D. Costa, J. R. Penadés, G. Peltz, Y. Matias, J. Manyika, D. Hassabis, Y. Xu, P. Kohli, A. Pawlosky, A. Karthikesalingam, and V. Natarajan, Accelerating scientific discovery with Co-Scientist, Nature 655, 487 (2026).DOI
  23. [23]L. Mitchener, A. Yiu, B. Chang, M. Bourdenx, T. Nadolski, A. Sulovari, E. C. Landsness, D. L. Barabási, S. Narayanan, N. Evans, S. Reddy, M. Foiani, A. Kamal, L. P. Shriver, F. Cao, A. T. Wassie, J. M. Laurent, E. Melville-Green, M. Caldas, A. Bou, K. F. Roberts, S. Zagorac, T. C. Orr, M. E. Orr, K. J. Zwedzaryk, A. E. Ghareeb, L. McCoy, B. Gomes, E. A. Ashley, K. E. Duff, T. Buonassisi, T. Rainforth, R. J. Bateman, M. Skarlinski, G. Rodrigues, M. M. Hinks, and A. D. White, Kosmos: An AI scientist for autonomous discovery (2025), arXiv:2511.02824 [cs.AI].
  24. [24]K. M. Jablonka, Q. Ai, A. Al-Feghali, S. Badhwar, J. D. Bocarsly, A. M. Bran, S. Bringuier, L. C. Brinson, K. Choudhary, D. Circi, S. Cox, W. A. de Jong, M. L. Evans, N. Gastellu, J. Genzling, M. V. Gil, A. K. Gupta, Z. Hong, A. Imran, S. Kruschwitz, A. Labarre, J. Lála, T. Liu, S. Ma, S. Majumdar, G. W. Merz, N. Moitessier, E. Moubarak, B. Mourinõ, B. Pelkie, M. Pieler, M. C. Ramos, B. Ranković, S. G. Rodrigues, J. N. Sanders, P. Schwaller, M. Schwarting, J. Shi, B. Smit, B. E. Smith, J. Van Herck, C. Völker, L. Ward, S. Warren, B. Weiser, S. Zhang, X. Zhang, G. A. Zia, A. Scourtas, K. J. Schmidt, I. Foster, A. D. White, and B. Blaiszik, 14 examples of how LLMs can transform materials science and chemistry: a reflection on a large language model hackathon, Digit. Discov. 2, 1233 (2023).
  25. [25]M. C. Ramos, C. J. Collison, and A. D. White, A review of large language models and autonomous agents in chemistry, Chem. Sci. 16, 2514 (2025).arxiv.org/abs/2407.01603
  26. [26]H. Zhang, Y. Li, W. Huang, Z. Hou, Y. Song, X. Liu, F. Effat, J. Jiang, S. Wu, Q. Ding, I. Takahara, L. R. MacGillivray, T. Mizoguchi, T. Yu, L. Liao, Y. Luo, Y. Rong, J. Li, Y. Diao, H. Ji, and B. Liu, Towards agentic intelligence for materials science (2026), arXiv:2602.00169 [cond-mat.mtrl-sci].arxiv.org/abs/2602.00169
  27. [27]L. S. Veiga, D. Zhang, T. Lu, Y. Lu, and H. Li, From prediction to realization: large language models and AI agents for inorganic materials discovery, Chem. Synth. 6, 66 (2026).DOI
  28. [28]A. Mirza, N. Alampara, S. Kunchapu, M. Ríos-García, B. Emokeba, A. Krishnan, T. Gupta, M. Schilling-Wilhelmi, M. Okereke, A. Aneesh, M. Asgari, J. Eberhardt, A. M. Elahi, H. M. Elbehery, M. V. Gil, C. Glaubitz, M. Greiner, C. T. Holick, T. Hoffmann, A. Ibrahim, L. C. Klepsch, Y. Köster, F. A. Kreth, J. Meyer, S. Miret, J. M. Peschel, M. Ringleb, N. C. Roesner, J. Schreiber, U. S. Schubert, L. M. Stafast, A. D. D. Wonnanke, M. Pieler, P. Schwaller, and K. M. Jablonka, A framework for evaluating the chemical knowledge and reasoning abilities of large language models against the expertise of chemists, Nat. Chem. 17, 1027 (2025).DOI
  29. [29]Anthropic, Formalizing Fermat’s Last Theorem, https://www.anthropic.com/research/formalizing-fermats-last-theorem (2026), published 4 September 2026.
  30. [30]OpenAI, On the Navier–Stokes Millennium Prize Problem, https://openai.com/index/navier-stokes-solution/ (2026), published 8 September 2026.
  31. [31]K. Kakaes, AI has solved one of math’s $1 million Millennium Prize problems, Quanta Magazine (2026), published 8 September 2026, https://www.quantamagazine.org/ai-has-solved-one-of-maths-1-million-millennium-prize-problems/
  32. [32]C. J. Bartel, Review of computational approaches to predict the thermodynamic stability of inorganic solids, J. Mater. Sci. 57, 10475 (2022).DOI
  33. [33]W. Sun, S. T. Dacek, S. P. Ong, G. Hautier, A. Jain, W. D. Richards, A. C. Gamst, K. A. Persson, and G. Ceder, The thermodynamic scale of inorganic crystalline metastability, Sci. Adv. 2, e1600225 (2016).
  34. [34]S. G. H. Kumar, Y. Zhou, A. Wang, J. Valdés-Hernández, T. W. Ko, N. Yue, O. Leng, H. Xu, C. Crebolder, A. Aspuru-Guzik, and V. Bernales, El Agente Sólido: A new age(nt) for solid state simulations (2026), arXiv:2602.17886 [cond-mat.mtrl-sci].
  35. [35]Z. Wang, H. Huang, H. Zhao, C. Xu, S. Zhu, J. Janssen, and V. Viswanathan, DREAMS: Density functional theory based research engine for agentic materials simulation (2025), arXiv:2507.14267 [cs.AI].
  36. [36]P. Yang, Z. Zhang, Y. Li, X. Wang, Y. Deng, Y. Lu, B. Tang, Z. Liu, and B. An, AutoDFT: A closed-loop multi-agent framework for autonomous DFT calculations (2026), arXiv:2605.26179 [cond-mat.mtrl-sci].arxiv.org/abs/2605.26179
  37. [37]A. Ghafarollahi and M. J. Buehler, Automating alloy design and discovery with physics-aware multimodal multi-agent AI, Proc. Natl. Acad. Sci. U.S.A. 122, e2414074122 (2025).DOI
  38. [38]A. Ghafarollahi and M. J. Buehler, Autonomous in-silico inorganic materials discovery via multi-agent physics-aware scientific reasoning, npj Comput. Mater. 12, 357 (2026).DOI
  39. [39]I. Takahara, T. Mizoguchi, and B. Liu, Accelerated inorganic materials design with generative AI agents, Cell Rep. Phys. Sci. 6, 103019 (2025).DOI
  40. [40]R. Nduma, H. Park, and A. Walsh, Crystalyses: a multi-tool agent for materials design (2025), arXiv:2512.00977 [cond-mat.mtrl-sci].arxiv.org/abs/2512.00977
  41. [41]B. Deng, B. Li, M. Cox, H. Chun, J. Nam, A. Lyssenko, S. Edamadaka, J. Ruza, X. Du, N. Segal, J. Diaz Sanchez, M. Xie, T. Perez, Y. Yao, M. Steiner, S. Majumdar, C. B. Musgrave, III, A. Chandra, A. Patra, D. Hohl, C. W. Coley, J. Li, and R. Gómez-Bombarelli, Harnessing AtomisticSkills for agentic atomistic research (2026), arXiv:2605.24002 [physics.chem-ph].arxiv.org/abs/2605.24002
  42. [42]Anthropic, Introducing the Model Context Protocol, https://www.anthropic.com/news/model-context-protocol (2024), specification at https://modelcontextprotocol.io.
  43. [43]H. Pan, R. Chard, R. Mello, C. Grams, T. He, A. Brace, O. P. Skelly, W. Engler, H. Holbrook, S. Y. Oh, M. Gonthier, M. Papka, B. Blaiszik, K. Chard, and I. Foster, Experiences with Model Context Protocol servers for science and high performance computing (2025), arXiv:2508.18489 [cs.DC].arxiv.org/abs/2508.18489
  44. [44]Materials Project, Model Context Protocol server of the Materials Project, https://docs.materialsproject.org/services/ml-and-ai-applications/mcp (2025), accessed 22 October 2026.
  45. [45]O. Kononova, H. Huo, T. He, Z. Rong, T. Botari, W. Sun, V. Tshitoyan, and G. Ceder, Text-mined dataset of inorganic materials synthesis recipes, Sci. Data 6, 203 (2019).DOI
  46. [46]T. He, H. Huo, C. J. Bartel, W. G. Krause, and G. Ceder, Precursor recommendation for inorganic synthesis by machine learning materials similarity from scientific literature, Sci. Adv. 9, eadg8180 (2023).DOI
  47. [47]S. Kim, Y. Jung, and J. Schrier, Large language models for inorganic synthesis predictions, J. Am. Chem. Soc. 146, 19654 (2024).DOI
  48. [48]T. Prein, E. Pan, J. Jehkul, S. Weinmann, E. A. Olivetti, and J. L. M. Rupp, Language models enable data-augmented synthesis planning for inorganic materials (2025), arXiv:2506.12557 [cond-mat.mtrl-sci].arxiv.org/abs/2506.12557
  49. [49]Z. Song, S. Lu, M. Ju, Q. Zhou, and J. Wang, Accurate prediction of synthesizability and precursors of 3D crystal structures via large language models, Nat. Commun. 16, 6530 (2025).DOI
  50. [50]D. W. Jeon, D. H. Kim, T. Jeon, J. Y. Na, H. Kim, H. Cho, D. S. Jung, J. Li, J.-S. Park, and S. B. Cho, Closed-loop solid-state synthesis planning for materials discovery with large language models, Adv. Mater. 38, e74502 (2026).DOI
  51. [51]M. Aykol, V. I. Hegde, L. Hung, S. Suram, P. Herring, C. Wolverton, and J. S. Hummelshøj, Network analysis of synthesizable materials discovery, Nat. Commun. 10, 2018 (2019).DOI
  52. [52]M. J. McDermott, S. S. Dwaraknath, and K. A. Persson, A graph-based network for predicting chemical reaction pathways in solid-state materials synthesis, Nat. Commun. 12, 3097 (2021).DOI
  53. [53]M. Aykol, J. H. Montoya, and J. Hummelshøj, Rational solid-state synthesis routes for inorganic materials, J. Am. Chem. Soc. 143, 9244 (2021).DOI
  54. [54]M. J. McDermott, B. C. McBride, C. E. Regier, G. T. Tran, Y. Chen, A. A. Corrao, M. C. Gallant, G. E. Kamm, C. J. Bartel, K. W. Chapman, P. G. Khalifah, G. Ceder, J. R. Neilson, and K. A. Persson, Assessing thermodynamic selectivity of solid-state reactions for the predictive synthesis of inorganic materials, ACS Cent. Sci. 9, 1957 (2023).DOI
  55. [55]J. Chen, S. R. Cross, L. J. Miara, J.-J. Cho, Y. Wang, and W. Sun, Navigating phase diagram complexity to guide robotic inorganic materials synthesis, Nat. Synth. 3, 606 (2024).DOI
  56. [56]N. J. Szymanski, P. Nevatia, C. J. Bartel, Y. Zeng, and G. Ceder, Autonomous and dynamic precursor selection for solid-state materials synthesis, Nat. Commun. 14, 6956 (2023).DOI
  57. [57]T. F. T. Cerqueira, A. Sanna, and M. A. L. Marques, Sampling the materials space for conventional superconducting compounds, Adv. Mater. 36, 2307085 (2024).DOI
  58. [58]F. Tran and P. Blaha, Accurate band gaps of semiconductors and insulators with a semilocal exchange-correlation potential, Phys. Rev. Lett. 102, 226401 (2009).pubmed.ncbi.nlm.nih.gov/19658882
  59. [59]P. Borldo, T. Aull, A. W. Huran, F. Tran, M. A. L. Marques, and S. Botti, Large-scale benchmark of exchange-correlation functionals for the determination of electronic band gaps of solids, J. Chem. Theory Comput. 15, 5069 (2019).DOI
  60. [60]D. K. Wilson, B. Sapparov, and S. Bobev, Synthesis, crystal structures and properties of the Zintl phases Sr₂ZnP₂, Sr₂ZnAs₂, A₂ZnSb₂ and A₂ZnBi₂ (A = Sr and Eu), Z. Anorg. Allg. Chem. 637, 2018 (2011).DOI
  61. [61]W. Shockley and H. J. Queisser, Detailed balance limit of efficiency of p-n junction solar cells, J. Appl. Phys. 32, 510 (1961).
  62. [62]S. Rühle, Tabulated values of the Shockley–Queisser limit for single junction solar cells, Sol. Energy 130, 139 (2016).DOI
  63. [63]M. F. Mansuetto, P. M. Keane, and J. A. Ibers, Synthesis, structure, and conductivity of the new group IV chalcogenides, KCuZrQ₃ (Q = S, Se, Te), J. Solid State Chem. 101, 257 (1992).DOI
  64. [64]S. Niu, B. Zhao, K. Ye, E. Bianco, J. Zhou, M. E. McConney, C. Settens, R. Haiges, R. Jaramillo, and J. Ravichandran, Crystal growth and structural analysis of perovskite chalcogenide BaZrS₃ and Ruddlesden–Popper phase Ba₃Zr₂S₇, J. Mater. Res. 34, 3819 (2019).DOI
  65. [65]P. Kayastha, D. Tiwari, A. Holland, O. S. Hutter, K. Durose, L. D. Whalley, and G. Longo, High-temperature equilibrium of 3D and 2D chalcogenide perovskites, Sol. RRL 7, 2201078 (2023).
  66. [66]A. Mesbah, S. Lebègue, J. M. Klingsporn, W. Stojko, R. P. Van Duyne, and J. A. Ibers, Synthesis, crystal structure, and optical properties of Ba₂Cu₂ThS₅, and electronic structures of Ba₂Cu₂ThS₅ and Ba₂Cu₂US₅, J. Solid State Chem. 200, 349 (2013).DOI
  67. [67]T. Nitta, K. Nagase, and S. Hayakawa, Formation, microstructure, and properties of barium zirconium sulfide ceramics, J. Am. Ceram. Soc. 53, 601 (1970).DOI
  68. [68]E. W. Staley, T. Arbaugh, M. Pekala, A. New, C. D. Stiles, N. Q. Le, G. Bassen, W. Bunstine, and T. McQueen, Coupling language models with physics-based simulation for synthesis of inorganic materials (2026), arXiv:2606.00315 [cs.AI].arxiv.org/abs/2606.00315
  69. [69]J. Schlesinger, S. Hjaltason, N. J. Szymanski, and C. J. Bartel, Thermodynamic assessment of machine learning models for solid-state synthesis prediction, Mater. Horiz. 10.1039/D6MH00209A (2026), arXiv:2602.04075 [cond-mat.mtrl-sci].arxiv.org/abs/2602.04075
  70. [70]C. W. Andersen, R. Armiento, E. Blokhin, G. J. Conduit, S. Dwaraknath, M. L. Evans, Á. Fekete, A. Gopakumar, S. Gražulis, A. Merkys, F. Mohamed, C. Oses, G. Pizzi, G.-M. Rignanese, M. Scheidgen, L. Talirz, C. Toher, D. Winston, R. Aversa, K. Choudhary, P. Colinet, S. Curtarolo, D. Di Stefano, C. Draxl, S. Er, M. Esters, M. Fornari, M. Giantomassi, M. Govoni, G. Hautier, V. Hegde, M. K. Horton, P. Huck, G. Huhs, J. Hummels, A. Kariryaa, B. Kozinsky, S. Kumbhar, M. Liu, N. Marzari, A. J. Morris, A. A. Mostofi, K. A. Persson, G. Petretto, T. Purcell, F. Ricci, F. Rose, M. Scheffler, D. Speckhard, M. Uhrin, A. Vaitkus, P. Villars, D. Waroquier, C. Wolverton, M. Wu, and X. Yang, OPTIMADE, an API for exchanging materials data, Sci. Data 8, 217 (2021).
  71. [71]N. J. Szymanski, B. Rendy, Y. Fei, R. E. Kumar, T. He, D. Milsted, M. J. McDermott, M. Gallant, E. D. Cubuk, A. Merchant, H. Kim, A. Jain, C. J. Bartel, K. Persson, Y. Zeng, and G. Ceder, An autonomous laboratory for the accelerated synthesis of novel materials, Nature 624, 86 (2023).DOI
  72. [72]G. Tom, S. P. Schmid, S. G. Baird, Y. Cao, K. Darvish, H. Hao, S. Lo, S. Pablo-García, E. M. Rajaoṇson, M. Skreta, N. Yoshikawa, S. Corapi, G. D. Akcoc, F. Strieth-Kalthoff, M. Seifrid, and A. Aspuru-Guzik, Self-driving laboratories for chemistry and materials science, Chem. Rev. 124, 9633 (2024).DOI

Paper details

vlnc:W_e5qx2zdg
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 11

downloads · 1

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents