Periodic Reporting for period 4 - MuDiLingo (A Multiscale Dislocation Language for Data-Driven Materials Science)
Periodo di rendicontazione: 2022-07-01 al 2023-10-31
Plasticity is mainly carried by dislocations – line-like defects in the crystal lattice that are responsible for a large range of mechanical properties of crystalline solids. Controlling the evolution of dislocations interacting both among themselves and with the other microstructures allows tailoring material behaviors on the micro and nanoscale. This is essential for microscale devices and rational design approaches towards next generation materials with superior properties.
For nearly a century, materials scientists have been seeking to understand how dislocation systems evolve. In-situ microscopy now reveals complex dislocation networks in great detail. However, due to the lack of a sufficiently versatile and general methodology for extracting, assembling and compressing dislocation-related information the analysis of such data often stays at the level of watching “images” to identify mechanisms or structures. Simulations are increasingly capable of predicting the evolution of dislocations in full detail.
Yet, a direct comparison, automated analysis or even data transfer between small scale plasticity experiments and simulations is impossible, and a large amount of data cannot be reused.
The vision of MuDiLingo is to develop and establish for the first time a Unifying Multiscale Language of Dislocation Microstructures. Bearing analogy to audio data conversion into MP3, this description of dislocations uses statistical methods to determine data compression, while preserving the relevant physics. It allows for a completely new type of high-throughput data mining and analysis, tailored to the specifics of systems of dislocations. This revolutionary approach will link different models and experiments on different length scales thereby guaranteeing true interoperability of simulations and experiments. The application to technologically relevant materials will answer fundamental scientific questions and guide towards the design of superior materials in the spirit of ICME.
For nearly a century, materials scientists have been seeking to understand how dislocation systems evolve. In-situ microscopy now reveals complex dislocation networks in great detail. However, due to the lack of a sufficiently versatile and general methodology for extracting, assembling and compressing dislocation-related information the analysis of such data often stays at the level of watching “images” to identify mechanisms or structures. Simulations are increasingly capable of predicting the evolution of dislocations in full detail.
Yet, a direct comparison, automated analysis or even data transfer between small scale plasticity experiments and simulations is impossible, and a large amount of data cannot be reused.
The vision of MuDiLingo is to develop and establish for the first time a Unifying Multiscale Language of Dislocation Microstructures. Bearing analogy to audio data conversion into MP3, this description of dislocations uses statistical methods to determine data compression, while preserving the relevant physics. It allows for a completely new type of high-throughput data mining and analysis, tailored to the specifics of systems of dislocations. This revolutionary approach will link different models and experiments on different length scales thereby guaranteeing true interoperability of simulations and experiments. The application to technologically relevant materials will answer fundamental scientific questions and guide towards the design of superior materials in the spirit of ICME.
==General Work Program==
1. learning python/C++
2. development of an elastic eigenstrain solver
3. creating of initial dislocation structures from DDD and MD
4. coarse graining and analysis of dislocation structures
5. first steps towards "machine learning dislocations"
6. creation of “synthetic” TEM images for machine learning
7. learning how to apply Deep Learning to TEM images
8. First steps towards segmentation of generic TEM images of dislocations
9. Sample preparation, testing and in-situ microscopy
10. Many discussions with the experimentalists how to obtain TEM images that are suitable for Machine Learning
==Overview of main results==
- We were the first to use machine learning to statistically analyse and characterize dislocation microstructures.
- We developed DL models that are now able to perform binary segmentation tasks even for very near-by or even overlapping dislocations. Acquiring training data for supervised learning was very difficult, which is also why we put a focus on the creating of synthetic training data, e.g. the generation of artificial TEM images. We were the first to do this in the materials science community.
- We developed, together with the experimental partners, a data-mining method where we can extract stresses and strain energy densities from digitized and 3D-reconstructed dislocation structures.
- The D2C method was further developed and could then also be used to characterize experimental TEM data.
- We developed the Dislocation Ontology as a “universal language for dislocation microstructures”.
- MuDiLingo was one of the very early projects within this part of the materials science community that leveraged the potential of machine learning methods.
The work has been publish in more than 25 articles in peer-review journals, e.g. in "Modelling and Simulation in Materials Science and Engineering", "Machine learning: science and technology", "Acta Materialia", "Frontiers in Materials" and "Compuational Materials Science". There is still a considerable amount of work currently under review and in preparation for publication. The results were also presented at international conferences, e.g. at "Material Science and Engineering (MSE)", "Materials Research Society (MRS)", "IUTAM Symposium on Data-Driven Mechanics" and "IEEE International Conference on Nanotechnology". Results also were and are to be documented in the PhD theses of the respective MuDiLingo staff.
1. learning python/C++
2. development of an elastic eigenstrain solver
3. creating of initial dislocation structures from DDD and MD
4. coarse graining and analysis of dislocation structures
5. first steps towards "machine learning dislocations"
6. creation of “synthetic” TEM images for machine learning
7. learning how to apply Deep Learning to TEM images
8. First steps towards segmentation of generic TEM images of dislocations
9. Sample preparation, testing and in-situ microscopy
10. Many discussions with the experimentalists how to obtain TEM images that are suitable for Machine Learning
==Overview of main results==
- We were the first to use machine learning to statistically analyse and characterize dislocation microstructures.
- We developed DL models that are now able to perform binary segmentation tasks even for very near-by or even overlapping dislocations. Acquiring training data for supervised learning was very difficult, which is also why we put a focus on the creating of synthetic training data, e.g. the generation of artificial TEM images. We were the first to do this in the materials science community.
- We developed, together with the experimental partners, a data-mining method where we can extract stresses and strain energy densities from digitized and 3D-reconstructed dislocation structures.
- The D2C method was further developed and could then also be used to characterize experimental TEM data.
- We developed the Dislocation Ontology as a “universal language for dislocation microstructures”.
- MuDiLingo was one of the very early projects within this part of the materials science community that leveraged the potential of machine learning methods.
The work has been publish in more than 25 articles in peer-review journals, e.g. in "Modelling and Simulation in Materials Science and Engineering", "Machine learning: science and technology", "Acta Materialia", "Frontiers in Materials" and "Compuational Materials Science". There is still a considerable amount of work currently under review and in preparation for publication. The results were also presented at international conferences, e.g. at "Material Science and Engineering (MSE)", "Materials Research Society (MRS)", "IUTAM Symposium on Data-Driven Mechanics" and "IEEE International Conference on Nanotechnology". Results also were and are to be documented in the PhD theses of the respective MuDiLingo staff.
- We were the first to use machine learning to statistically analyse and characterize dislocation microstructures.
- We developed DL models that are now able to perform binary segmentation tasks even for very near-by or even overlapping dislocations. Acquiring training data for supervised learning was very difficult, which is also why we put a focus on the creating of synthetic training data, e.g. the generation of artificial TEM images. We were the first to do this in the materials science community.
- Our work summarized in the article “Data-mining of in-situ TEM experiments: On the dynamics of dislocations in CoCrFeMnNi alloys”, published in Acta Materialia (241)2022, significantly advanced the knowledge about the motion and interaction of dislocations in high-entropy alloys and in particular also showed a new way how to analyze such situations. In particular, we demonstrated that the interaction of dislocations with “obstacles” in high-entropy alloys can result in hardening or in softening and is additionally dependent on the number of passing dislocations. So far, it was not possible to quantitatively analyze in-situ TEM imaging of dislocation in such a detailed degree.
- Our work summarized in the article "“Machine Learning-Based Classification of Dislocation Microstructures” in Frontiers in Materials, (6)2019, paved the way towards machine learning-assisted theory development for dislocation dynamics.
- We have done groundbreaking work in developing such a Dislocation Ontology, which is now also being used in the German Research Data Infrastructure effort (NFDI) and to be used within European efforts (connecting to the EMMO/EMC).
- We developed DL models that are now able to perform binary segmentation tasks even for very near-by or even overlapping dislocations. Acquiring training data for supervised learning was very difficult, which is also why we put a focus on the creating of synthetic training data, e.g. the generation of artificial TEM images. We were the first to do this in the materials science community.
- Our work summarized in the article “Data-mining of in-situ TEM experiments: On the dynamics of dislocations in CoCrFeMnNi alloys”, published in Acta Materialia (241)2022, significantly advanced the knowledge about the motion and interaction of dislocations in high-entropy alloys and in particular also showed a new way how to analyze such situations. In particular, we demonstrated that the interaction of dislocations with “obstacles” in high-entropy alloys can result in hardening or in softening and is additionally dependent on the number of passing dislocations. So far, it was not possible to quantitatively analyze in-situ TEM imaging of dislocation in such a detailed degree.
- Our work summarized in the article "“Machine Learning-Based Classification of Dislocation Microstructures” in Frontiers in Materials, (6)2019, paved the way towards machine learning-assisted theory development for dislocation dynamics.
- We have done groundbreaking work in developing such a Dislocation Ontology, which is now also being used in the German Research Data Infrastructure effort (NFDI) and to be used within European efforts (connecting to the EMMO/EMC).