main concepts scientists sections
SimplePopularAdvanced

From Computer Prediction to Lab Synthesis: How to Find 25 Stable Materials Out of 894 ✛ ✎ ✦

Original: "Predicting Novel Stable Materials for Experimental Synthesis"
· Yuqi An, Sihong Zhu, Joseph Montoya, Xingyu Guo, Zhenbin Wang
arXiv:2607.01713 · 2026-07-02 · arXiv non-exclusive · independent analysis · 4 min
Materials Computational Physics
Hierarchical screening based on thermodynamics and dynamics leaves only a few percent of predicted structures truly ready for synthesis.

Context

Back in his famous lecture, Richard Feynman talked about simulating physics on computers, and today we stand on the threshold of the era of predictive materials science. The rapid development of computational chemistry and machine learning has created the illusion that a new material can be discovered simply by pressing a button. Indeed, algorithms find millions of stable atomic combinations, but experimentalists face a harsh reality: obtaining the predicted compound in the lab is incredibly difficult. The reason is that calculations often rely on simplified models and incomplete data, ignoring the second law of thermodynamics, which requires considering all possible structures. Without dynamic verification—whether the crystal will maintain its shape when heated—the prediction remains a gamble. This gap between computer prediction and real synthesis has become the main bottleneck in materials science.

Methods

The authors built a three-level sieve, reminiscent of special forces selection: candidates undergo tests of increasing difficulty. First, from the CAMD database, 603 unique structures were selected and their stability was recalculated using numerical simulation based on the PBE functional (electron density approximation). Here, the key tool is the energy above the convex hull, showing how 'unfavorable' the compound is compared to known ones. The second stage—dynamic stability: using the universal machine learning potential MACE, phonon spectra were calculated (imaginary modes are a sure sign of instability) and molecular dynamics was run on a nanosecond timescale at 1000 K. The final filter is recalculation with the more accurate SCAN functional, which halves the errors in formation energies.

Results

The very first screening was staggering: out of 603 unique materials, only 298 (49%) remained thermodynamically stable on the full phase diagram. The reason lies in competing phases—the addition of previously unknown vanadium bromide VBr3 shifted VBr5 from stable to metastable, raising its energy by 29 meV/atom. Dynamic tests, accounting for entropic effects, shrank the list to 166: 25 structures failed harmonic oscillations, and 107 showed instability in molecular dynamics. Notably, a third of materials behaved inconsistently—phonon stability did not guarantee stability under heating, and vice versa. SCAN refinement left 109 finalists. Finally, considering decomposition enthalpy and chemical space completeness highlighted 25 candidates with a score above 0.80—true diamonds in a pile of sand.

Implications

This work shifts the paradigm: from now on, any predicted material must undergo not only thermodynamic but also dynamic scrutiny. The proposed pipeline dramatically increases confidence in computer design, bringing it closer to the needs of experimentalists. For synthetic chemists, this means a clear list of the 25 most promising targets—instead of thousands of speculative options. For fundamental science, the protocol underscores the importance of the Pauli exclusion principle and the Schrödinger equation, which underpin DFT calculations, and demonstrates how quantum thermodynamics manifests at the nanoscale. Potential applications range from superconductors to quantum computers and quantum information transmission, where crystal lattice stability is critical. The reliability of these calculations rests on the mathematical heart of quantum mechanics, discovered by Erwin Schrödinger. Experimental verification of such structures will require advanced spectroscopy methods.

Future development

The next twist is incorporating variable cell shape in molecular dynamics to catch phase transitions requiring volume changes. The development of universal ML potentials, for instance, accounting for charge states, will make predictions even more accurate and allow tackling ternary and quaternary systems. Integration of such a pipeline into automated laboratories, where a robotic synthesizer will receive a pre-validated list of compounds, is not far off.

Impact

The methodology will directly impact the creation of new batteries, catalysts, and quantum devices, where every extra millielectronvolt of energy determines the material's fate.

Next steps

The immediate task is the synthesis of the 25 proposed compounds by leading experimental groups, followed by structural validation. Simultaneously, expanding databases of competing phases and training ML potentials on a broader class of chemical elements are needed.

Key open problems

The study strikes at the heart of an unsolved problem in condensed matter physics—why most ab initio-stable phases do not form in experiments. It also raises questions about the limits of the PBE functional and the role of dynamic chaos in disrupting ordered structures.

The knowledge machine's view

This section is addressed to the author of the paper and their colleagues. Neighbouring works were found by meaning-based search across our archive, not by word overlap. We do not review: we show the field around the work.

What the work does

The authors built a three-tier filter to select stable materials from the CAMD database: first, they recalculated the thermodynamic stability of 603 structures using PBE, then checked dynamic stability via phonon spectra and molecular dynamics with the MACE potential, and finally refined the energies with the SCAN functional. As a result, out of the 894 initial candidates, 25 most promising ones for synthesis remained.

Where its strength is

The strength of the work lies in combining thermodynamic and dynamic stability criteria, which sharply reduces the number of false-positive predictions and gives experimentalists a concrete shortlist of targets.

Where it could go next

Worth trying

  • Check how well the stability predictions correlate with experimental synthesis data from the literature.
  • Try incorporating kinetic accessibility (reaction barriers) into the pipeline for the final candidates.
  • Use active learning to iteratively refine the list as new experimental results become available.

Nearby works in our archive

The closest work in the archive sits at 0.63 by meaning — the usual level for our corpus.

Whole graph →

Out of 894 initial materials, only 25 passed all tests: that's less than 3%, comparable to the percentage of gold in seawater. Nature seems to deliberately hide the most stable forms.

The idea of computerized material prediction echoes Neal Stephenson's novel 'The Diamond Age,' where nanotechnology creates matter from digital blueprints. However, our current level is closer to Greg Egan's 'Miranda'—we still painstakingly test each virtual structure for viability.

E_{\text{hull}} = E_{A} - \sum_i x_i E_i^{\text{stable}}
Difference between the total energy of compound A and the energy of the mixture of the most stable competing phases of the same composition. A zero value indicates thermodynamic stability.

Key numbers

  • initial materials in CAMD: 894
  • unique structures after cleaning: 603
  • thermodynamically stable by PBE: 298 (49.4%)
  • dynamically stable: 166
  • stable by SCAN: 109
  • high-priority for synthesis: 25
Scientists
David DeutschErwin SchrödingerJohn Stewart BellLouis de BroglieLov GroverPaul Benioff
Concepts
machine learning materials materials discovery thermodynamic stability ai drug discovery candidate training stiffness predictive learning neural networks training data drug design dataset protein language model synthetic data synthesis thermal conductivity prediction benzene machine learned interatomic potential protein design stable state predictive model