Periodic Reporting for period 4 - Materials 4.0 (Advancing materials design by high-accuracy finite-temperature first principles calculations accelerated by machine learning potentials)
Período documentado: 2025-07-01 hasta 2025-12-31
Modern materials design relies on phase diagrams: “maps” that tell us which crystal structure or alloy phase is stable at a given temperature and composition. These maps are essential for selecting materials for turbines, engines, power plants, batteries, and many other technologies, because they determine which phases form during processing and which remain stable in service.
A key difficulty is that many widely used computational materials databases and workflows still focus on the 0 K limit. 0 K (“zero Kelvin”) is absolute zero, i.e. –273.15 °C, the theoretically lowest possible temperature. At 0 K, atoms are assumed to be located in their ideal positions without vibrations, making calculations much easier. However, real materials are manufactured and operate at hundreds or thousands of degrees, where thermal vibrations (often strongly anharmonic), electronic excitations, and defects can decisively change stability and properties.
The project tackled this gap using ab initio simulations. “Ab initio” (Latin for “from the beginning”) means simulations that start from the fundamental laws of quantum mechanics to compute material properties without fitting to experimental data. In practice, this usually means density-functional theory (DFT) as the baseline quantum-mechanical method. Ab initio simulations can be highly predictive, but they become costly when one needs to describe realistic temperature effects, large supercells, chemical disorder (alloys), and long time scales.
The overall objective of Materials 4.0 was therefore to develop and validate a practical, high-accuracy finite-temperature simulation framework that enables predictive thermodynamics for complex materials. Inspired by the “Industry 4.0” idea of connecting processes through data, Materials 4.0 coupled ab initio physics with data-driven acceleration: machine-learning interatomic potentials provide fast sampling of thermal motion, while carefully designed “upsampling” steps (free-energy perturbation) recover ab initio accuracy for free energies and derived properties.
By the end of the project, Materials 4.0 delivered a validated end-to-end methodology for computing finite-temperature thermophysical properties (e.g. heat capacity, thermal expansion, elastic response) up to very high temperatures, including regimes where standard low-temperature approximations fail. The approach was demonstrated and benchmarked on various elements and extended to chemically complex alloys (including high-entropy alloys). In addition, the workflow was pushed beyond standard DFT accuracy for selected phase-transition problems by combining machine learning with upsampling across exchange–correlation functionals, and it enabled large-scale defect simulations (e.g. superdislocations) with physically informed active learning potentials. Overall, the project established a reliable platform for more accurate and efficient phase-stability predictions and simulation-driven materials development at realistic temperatures.
A key difficulty is that many widely used computational materials databases and workflows still focus on the 0 K limit. 0 K (“zero Kelvin”) is absolute zero, i.e. –273.15 °C, the theoretically lowest possible temperature. At 0 K, atoms are assumed to be located in their ideal positions without vibrations, making calculations much easier. However, real materials are manufactured and operate at hundreds or thousands of degrees, where thermal vibrations (often strongly anharmonic), electronic excitations, and defects can decisively change stability and properties.
The project tackled this gap using ab initio simulations. “Ab initio” (Latin for “from the beginning”) means simulations that start from the fundamental laws of quantum mechanics to compute material properties without fitting to experimental data. In practice, this usually means density-functional theory (DFT) as the baseline quantum-mechanical method. Ab initio simulations can be highly predictive, but they become costly when one needs to describe realistic temperature effects, large supercells, chemical disorder (alloys), and long time scales.
The overall objective of Materials 4.0 was therefore to develop and validate a practical, high-accuracy finite-temperature simulation framework that enables predictive thermodynamics for complex materials. Inspired by the “Industry 4.0” idea of connecting processes through data, Materials 4.0 coupled ab initio physics with data-driven acceleration: machine-learning interatomic potentials provide fast sampling of thermal motion, while carefully designed “upsampling” steps (free-energy perturbation) recover ab initio accuracy for free energies and derived properties.
By the end of the project, Materials 4.0 delivered a validated end-to-end methodology for computing finite-temperature thermophysical properties (e.g. heat capacity, thermal expansion, elastic response) up to very high temperatures, including regimes where standard low-temperature approximations fail. The approach was demonstrated and benchmarked on various elements and extended to chemically complex alloys (including high-entropy alloys). In addition, the workflow was pushed beyond standard DFT accuracy for selected phase-transition problems by combining machine learning with upsampling across exchange–correlation functionals, and it enabled large-scale defect simulations (e.g. superdislocations) with physically informed active learning potentials. Overall, the project established a reliable platform for more accurate and efficient phase-stability predictions and simulation-driven materials development at realistic temperatures.
A central outcome was the development, implementation, and broad application of the direct upsampling framework. In this approach, machine-learning potentials are used to efficiently sample temperature-dependent atomic motion, and a rigorous free-energy correction step (“upsampling”) restores results to ab initio accuracy. This enables computation of heat capacities, thermal expansion, elastic response, and related quantities over a wide temperature range, where conventional low-temperature approximations break down.
The methodology was validated across multiple elemental systems (including high-melting refractory metals) and subsequently transferred to chemically complex alloys, such as refractory high-entropy alloys. The results show that high-temperature excitations (anharmonic vibrations, electronic excitations, and their coupling) can make a decisive contribution and must be treated consistently to achieve quantitative agreement with experimental reference data.
In the later project phase, the workflow was pushed beyond standard DFT for selected problems by performing upsampling in the exchange–correlation space. This enabled a systematic comparison of different levels of electronic-structure theory for finite-temperature phase transitions and, for silica polymorphs, demonstrated that only the fifth rung (random-phase approximation) on the so-called Jacob’s ladder yields an accurate transition temperature when entropy differences are slight.
Beyond thermodynamics, the project developed physically informed active learning strategies to build transferable machine-learning potentials for complex defect processes. This enabled large-scale atomistic simulations of superdislocations in Ni3Al, thereby providing microscopic insight into the origin of the yield-stress anomaly in Ni-based superalloys.
The project’s results were widely disseminated through peer-reviewed publications, conference and invited presentations, and open research outputs. Overall, 48 journal articles were published (with additional manuscripts under review), two software packages were released, and more than 15 datasets were made publicly available, supporting transparency, benchmarking, and reuse. The scientific visibility also facilitated knowledge transfer and external uptake, including an industry-funded collaboration.
The methodology was validated across multiple elemental systems (including high-melting refractory metals) and subsequently transferred to chemically complex alloys, such as refractory high-entropy alloys. The results show that high-temperature excitations (anharmonic vibrations, electronic excitations, and their coupling) can make a decisive contribution and must be treated consistently to achieve quantitative agreement with experimental reference data.
In the later project phase, the workflow was pushed beyond standard DFT for selected problems by performing upsampling in the exchange–correlation space. This enabled a systematic comparison of different levels of electronic-structure theory for finite-temperature phase transitions and, for silica polymorphs, demonstrated that only the fifth rung (random-phase approximation) on the so-called Jacob’s ladder yields an accurate transition temperature when entropy differences are slight.
Beyond thermodynamics, the project developed physically informed active learning strategies to build transferable machine-learning potentials for complex defect processes. This enabled large-scale atomistic simulations of superdislocations in Ni3Al, thereby providing microscopic insight into the origin of the yield-stress anomaly in Ni-based superalloys.
The project’s results were widely disseminated through peer-reviewed publications, conference and invited presentations, and open research outputs. Overall, 48 journal articles were published (with additional manuscripts under review), two software packages were released, and more than 15 datasets were made publicly available, supporting transparency, benchmarking, and reuse. The scientific visibility also facilitated knowledge transfer and external uptake, including an industry-funded collaboration.
Before Materials 4.0 predictive finite-temperature ab initio thermodynamics was often limited by two bottlenecks: (1) insufficient physics in standard low-temperature approximations (e.g. quasi-harmonic treatments that miss strong anharmonicity and electronic effects at high temperature), and (2) insufficient scale, since direct ab initio sampling is too expensive for the large supercells needed for alloys, disorder, and defects.
Materials 4.0 advanced the state of the art by introducing and validating a workflow that is both physically complete and computationally scalable: machine-learning potentials enable extensive sampling at realistic temperatures, while upsampling provides a controlled route back to ab initio free energies. This allows high-temperature property predictions up to near the melting point, including cases where multiple excitation mechanisms contribute comparably and must be treated consistently.
Beyond this core capability, the project achieved two additional step-changes:
1. Systematic finite-temperature accuracy upgrades beyond standard DFT through upsampling in the exchange–correlation space accelerated by machine learning, demonstrated on a demanding silica polymorph transition where minor errors have outsized consequences. 2. Large-scale defect simulations with reliability, enabled by physically informed active-learning potentials, provide mechanistic insight, for example, into superdislocations and high-temperature deformation phenomena relevant to Ni-based superalloys.
Together, these results establish a durable platform for future predictive materials modeling: more reliable phase stability and thermophysical databases at finite temperatures, improved integration with thermodynamic assessments, and extension to additional phenomena in which temperature, disorder, and defects interact.
Materials 4.0 advanced the state of the art by introducing and validating a workflow that is both physically complete and computationally scalable: machine-learning potentials enable extensive sampling at realistic temperatures, while upsampling provides a controlled route back to ab initio free energies. This allows high-temperature property predictions up to near the melting point, including cases where multiple excitation mechanisms contribute comparably and must be treated consistently.
Beyond this core capability, the project achieved two additional step-changes:
1. Systematic finite-temperature accuracy upgrades beyond standard DFT through upsampling in the exchange–correlation space accelerated by machine learning, demonstrated on a demanding silica polymorph transition where minor errors have outsized consequences. 2. Large-scale defect simulations with reliability, enabled by physically informed active-learning potentials, provide mechanistic insight, for example, into superdislocations and high-temperature deformation phenomena relevant to Ni-based superalloys.
Together, these results establish a durable platform for future predictive materials modeling: more reliable phase stability and thermophysical databases at finite temperatures, improved integration with thermodynamic assessments, and extension to additional phenomena in which temperature, disorder, and defects interact.