Rapid estimation of synthesizability windows of inorganic materials from first principles

This paper presents a high-throughput method combining machine-learned interatomic potentials and fitted reference energies to rapidly generate temperature- and pressure-dependent phase predominance diagrams, enabling the efficient estimation of synthesizability windows for inorganic materials while overcoming the computational limitations of conventional DFT-based phonon calculations.

The Gist

Imagine you are a chef trying to bake a perfect cake. You have a recipe (the chemical formula), but you don't know the right oven temperature or how much humidity to keep in the kitchen. If the oven is too hot, the cake burns; if it's too cold, it never rises. In the world of materials science, scientists are the chefs, and "inorganic materials" (like metals, oxides, and sulfides) are the cakes.

For a long time, scientists have had a way to predict if a cake could exist, but only if they were baking in a perfect, frozen world at absolute zero (0 Kelvin). This is like checking if the ingredients can physically fit together in a box without any heat. However, real life isn't frozen. Real synthesis happens in hot, pressurized ovens with gases flowing around. The old "frozen world" maps often failed to tell scientists the right temperature or gas pressure to actually make the material.

The Problem: The "Frozen Map" vs. The "Real Kitchen"
The paper argues that the old method is like using a map of a city in winter to navigate it in summer. It misses the melting snow and the open roads. Calculating the "summer map" (how materials behave at high heat) used to be incredibly slow and expensive, like trying to simulate every single molecule dancing in the oven. It took so much computer power that scientists couldn't do it for thousands of materials at once.

The Solution: A New, Fast "Weather Forecast" for Materials
The authors developed a new, fast workflow to create "Synthesizability Windows." Think of this as a dynamic weather forecast for your material. Instead of just saying "this cake exists," it tells you: "To bake this cake, you need an oven at 500°C with a specific amount of oxygen gas."

They did this by combining three tools:

1. The Blueprint (DFT): They used standard computer models to get the basic structure of the material.
2. The Correction (FERE): They realized their blueprints were slightly off, like a recipe that always calls for too much salt. They added a "tuning knob" (called Fitted Elemental-Phase Reference Energies) to adjust the numbers so they matched real-world experiments much better.
3. The Speedster (MLIP): This is the magic trick. Instead of calculating the heat and movement of atoms the slow, traditional way, they used a "Machine-Learned Interatomic Potential" (MLIP). Imagine this as a super-smart AI that has watched millions of atoms dance and can instantly guess how they will move and vibrate at high temperatures. This step, which used to take days, now takes just a few minutes.

What They Found
They tested this new method on four families of materials: Oxides (rust-like), Nitrides, Sulfides, and Phosphides. They also applied it to a massive, complex group of 48 different "Metal Phosphosulfide" systems (think of these as complicated multi-layered cakes).

Here are the key takeaways from their "kitchen experiments":

• Metastable Materials Come to Life: Some materials that looked "dead" or impossible in the frozen 0-Kelvin map actually come alive when you add heat. For example, a material called Cu3P looked unstable in the old maps, but the new "weather forecast" showed it has a perfect window of temperature and pressure where it thrives. This explains why chemists have been able to make it in the lab for years, even though the old math said they shouldn't be able to.
• The "False Negatives": Sometimes, the new map shows a material is stable, but the old experimental records don't list it. The authors suggest this might be because scientists spent years trying to force unstable materials to exist using tricky, non-standard methods. The new map suggests that the "easy" materials to make are actually the ones that have a natural, stable window.
• Phase Transitions: The method can predict when a material will change its "shape" (polymorph) as it gets hotter. For instance, a material might be square-shaped at low temperatures but turn into a rectangle at high temperatures. The new diagrams show exactly when this switch happens.
• Speed and Scale: They generated these detailed maps for over 1,000 different compounds. Because the MLIP tool is so fast, they can do this for almost any inorganic material without waiting weeks for a computer to finish the math.

The Bottom Line
This paper presents a new, fast, and accurate way to tell experimental scientists exactly how to cook their materials. By translating complex computer energy calculations into simple "Temperature vs. Gas Pressure" maps, they bridge the gap between theoretical predictions and the actual lab bench. It turns a guess-and-check process into a guided recipe, helping scientists discover and create new materials much faster.

Technical Summary

Technical Summary: Rapid Estimation of Synthesizability Windows of Inorganic Materials from First Principles

Problem Statement
Predicting the synthesizability conditions of inorganic materials remains a significant challenge, even under the assumption of thermodynamic equilibrium. Traditional approaches relying solely on zero-temperature convex stability hulls ($E_h$) fail to account for finite-temperature effects, such as vibrational entropy and heat capacity, which are critical for realistic synthesis conditions. Conversely, accurate phase diagram calculations at finite temperatures typically require computationally expensive phonon calculations based on Density Functional Theory (DFT) and extensive sampling for configurational entropy. These methods are often too slow for high-throughput screening. Furthermore, existing machine learning (ML) approaches based on text-mined synthesis recipes or heuristic descriptors often suffer from limited generalizability and a tendency to overestimate synthesis likelihood. There is a need for a physics-based, high-throughput workflow that bridges the gap between computational predictions and experimental synthesis parameters (temperature and gas partial pressures) without incurring the prohibitive cost of full DFT phonon calculations.

Methodology
The authors propose an accelerated workflow to generate phase predominance diagrams as a function of temperature and partial pressures of gaseous reactants. The methodology integrates three primary components:

1. Formation Enthalpy Correction (FERE): Zero-temperature total energies are calculated using DFT (PBEsol functional). To improve accuracy, the authors employ the Fitted Elemental-Phase Reference Energies (FERE) approach. This involves fitting corrections to the elemental reference phase energies to minimize the difference between calculated and experimentally measured formation enthalpies ($\Delta H_f$) for specific datasets (oxides, sulfides, and pnictides).
2. Rapid Vibrational Properties via MLIPs: Instead of performing costly DFT phonon calculations, the workflow utilizes the MatterSim machine-learned interatomic potential (MLIP) foundation model. This model is used to rapidly calculate phonon band structures and densities of states (DOS) for vibrational entropy ($S$) and heat capacity ($C_p$). The authors note that while MLIPs are semi-empirical, they achieve mean absolute errors (MAE) of only 15 J/(mol·K) for vibrational entropy and 3 J/(mol·K) for heat capacity compared to explicit DFT calculations.
3. Thermodynamic Integration: The calculated $\Delta H_f$, $S$, and $C_p$ are integrated using standard thermochemical relationships (via HSC Chemistry software) to compute the Gibbs free energy ($\Delta G$) as a function of temperature and gas partial pressures ($p_{O_2}, p_{N_2}, p_{S_2}, p_{P_2}$). The resulting phase predominance diagrams identify the thermodynamically stable phase at any given coordinate point.

Key Results
The study validates the workflow across binary and ternary systems:

• Binary Systems (Oxides, Nitrides, Sulfides, Phosphides): The method was tested on V-O, Ta-N, Sn-S, and Cu-P systems.
  • Accuracy: After FERE corrections, the MAE for formation enthalpy was reduced significantly (e.g., from 0.236 to 0.069 eV/atom for oxides). The errors in MLIP-calculated entropy and heat capacity were found to be negligible compared to formation enthalpy errors at temperatures below ~1500 K.
  • Comparison with Experiment: The predicted phase boundaries showed good agreement with experimental thermochemical data and reported synthesis conditions. The mean absolute error in phase boundary temperature was 270 K.
  • Metastability: The approach successfully identified stability windows for compounds that appear metastable at 0 K (e.g., $V_3O_5$, $V_6O_{11}$, $Cu_3P$), demonstrating that finite-temperature effects can stabilize materials above the convex hull.
• Ternary Systems (Metal Phosphosulfides): The workflow was scaled to 48 distinct metal phosphosulfide ternary systems, covering over 1000 ternary compounds and ~450 binaries.
  • High-Throughput Capability: The computational cost was drastically reduced, requiring only ~30 core-minutes for DFT relaxation and ~5 core-minutes for MLIP phonon calculations per material.
  • New Discoveries: The method identified 19 new materials lying on the 0 K stability hull. It also predicted thermodynamic stability windows for metastable compounds (e.g., $HfP_2S_6$) and identified temperature-driven phase transitions between polymorphs (e.g., in $GaPS_4$, $Li_3PS_4$, and $Na_2PS_3$).
  • Validation: For the Sn-P-S system, the predicted stability windows for $SnP_2S_6$ and $SnPS_3$ aligned remarkably well with literature synthesis conditions.

Significance and Claims
The paper claims to demonstrate a physics-based workflow that enables the high-throughput generation of phase predominance diagrams for ordered inorganic materials. The primary significance lies in:

1. Bridging the Gap: Translating computational energies (often above the stability hull) into intuitive, experimentally relevant synthesis conditions (temperature vs. partial pressure).
2. Efficiency: Circumventing the bottleneck of DFT phonon calculations by utilizing MLIPs, which provide sufficient accuracy for entropy and heat capacity without compromising the overall predictive power of the phase diagrams.
3. Predictive Capability: Showing that zero-temperature stability metrics alone are insufficient; finite-temperature modeling is essential to correctly identify synthesizable compounds, including those that are metastable at 0 K.
4. Scalability: Proving the method's applicability to complex multinary systems (ternary phosphosulfides) with two gas-phase reactants, facilitating the discovery and synthesis of computationally identified materials.

The authors conclude that this approach facilitates communication between theorists and experimentalists and supports closed-loop autonomous labs for materials discovery, while noting that future work could incorporate configurational entropy, defects, and kinetic effects.