25. Neural network approach to the reverse transformation of spin-wave theory

 

Shravani Chillal, shravani.chillal@helmholtz-berlin.de 

 We study novel magnetism exhibited in quantum magnets and materials with strongly correlated electrons. Our goal is to understand/predict the magnetic properties in these systems through experimental investigation in conjunction with theoretical and machine learning models.

What is the data science project’s research question? 

In this project, we would like to reproduce the magnetic Hamiltonian of a known two-dimensional quantum magnet BaNi2V2O8 (BNVO) using machine learning techniques. The Hamiltonian of BNVO has been previously determined by fitting the inelastic neutron scattering (INS) data to a known physical model that has been approximated to contain 5 Hamiltonian parameters, J1-J5. It is to be noted that such analysis often also considers constraints from further experimental results while determining the values in addition to the researcher’s intuition based on experience. In some magnets where quantum effects are dominant, any small perturbation to the Hamiltonian can radically affect the inelastic excitation spectra. Consequently, the traditional fitting approach becomes either very time-consuming or impossible due to the multiple spurions or artifacts in the scattering data. Therefore, in order to independently verify the Hamiltonian parameters of BNVO without additional input from other physical properties we would like to use machine learning algorithms.
In summary, we would like to achieve two goals using machine learning methods on simulated INS data:
1. Accurate removal of the systematic/non-systematic, reproducible/irreproducible artifacts in the experimental data.
2. Determination of the magnetic Hamiltonian parameters from the experimental data.

What data will be worked on? 

The experimental INS data is measured at the neutron facility ISIS (Didcot, United Kingdom). The theoretical training data sets are simulated using Matlab-based package spinW and are stored in HDF5 format which is accessible by many programs. These datasets contain values of the 4-dimensional function S(Qx, Qy, Qz,,ω) for a single crystal material and 2-dimensional S(|Q|,ω) for a powder sample of BNVO. While the former data provides a clear resolution and hence better predictions from the model, they are prone to insufficient computing power due to the large size (approx. 250 GB per measurement). On the other hand, the latter data enables only an approximate solution due to the lack of information. Hence, we would like to use an optimized method that can make use of both the datasets to predict the magnetic Hamiltonian parameters.
In this project you will work with the powder (~100,000 datasets) and single crystal (~250,000 cuts & ~3,000 4D-datasets) data simulated for different combinations of magnetic exchange interactions. All the experimental data will be cleaned to a high-level and will be provided in a HDF5 (or Tiff-16bit) format.

What tasks will this project involve?  

First objective:
1. Extract the magnetic exchange interactions from the powder data (2-dimensional data)
o Train simulated S(|Q|,ω) images to obtain a best matching pattern with respect to experimental data.
o Use different machine learning models to automate this process
o Deliver a Python/Matlab code, ideally end-to-end, that delivers the best matched input.
2. Identification and removal of artifacts in the experimental S(|Q|,ω)
o Categorize the sample-based and non-sample based artifacts.
o Use different machine learning models such as an Auto-encoder to automate this process.
o Combine with the first objective in an iterative way so that total S(|Q|,ω) can be trained.
Second objective:
3. Extend the learned model to include the single crystal data (4-dimensional) in determining the Hamiltonian.

What makes this project interesting to work on?  Use of machine learning algorithms for data analysis is new in neutron scattering science and, there are very few cases where it has been applied to INS data. Therefore, this project could significantly improve the quality, quantity and accuracy of the information that can be extracted from the data. A successful completion of this project will not only advance data analysis in our current and future projects, but also help reconfirm/revisit the findings of many other magnetic materials that are fully/partially understood. Hence, we expect to use this work in several publications.

What is the expected outcome?   Contribution to research paper, Contribution to software development

What infrastructure, programs and tools will be used? Can they be used remotely?   

• Python, Matlab
• 4 department server (Window & Linux) and central cluster computer (centrum GPU cluster in planning)
• GPU cards: Tesla: 2x M2075, K40m & V 100

* Remote work is possible

What skills are necessary for this project?  Data analytics / statistics, Data mining / Machine learning, Deep learning, Visualization

Is the data open source?  No

Interested candidates should be at PhD  level. Shravani Chillal is looking for 1 visiting scientist, working on the project together with the team.