AI-driven scientific data analysis is becoming an important part of research workflows that combine complex experiments with computational models. Topics at the 2019 APS/CNM Users Meeting at Argonne National Laboratory included machine learning for nanoscale materials modeling, automatic differentiation for coherent imaging, and the data demands associated with new research capabilities. These examples offer a useful look at how advanced analysis methods are applied when experimental data is difficult to interpret with conventional approaches alone.
The work described at the meeting focused on scientific research, not industrial automation deployments. However, the underlying ideas—connecting a physics-based model to measured data, automating demanding calculations, and choosing analysis methods suited to the problem—are relevant to technical teams assessing advanced imaging and data workflows.
AI-driven scientific data analysis starts with the measurement problem
Advanced analysis is most useful when it addresses a specific limitation in how data can be collected or interpreted. At the Advanced Photon Source (APS), researchers use X-rays to investigate material structure and behavior. The Center for Nanoscale Materials (CNM) supports research that includes nanoscale characterization and modeling. Both settings produce data whose meaning depends on the experimental geometry, the materials being studied, and the physical processes behind the signal.
One example is coherent diffraction imaging (CDI), which can produce images at nanometer resolution in a range of experimental arrangements. CDI reconstruction is a phase retrieval problem: the measurement provides the amplitude of an object’s Fourier transform, but the phase needed to reconstruct the object is not directly available. Recovering the image therefore requires solving a nonlinear optimization problem.
Automatic differentiation for phase retrieval
Gradient-based optimization is commonly used to solve phase retrieval problems. In a conventional implementation, developers may need to derive the gradient equations for each experimental setup. That work can be difficult and time-consuming, especially when the imaging scheme is complex. It can also make it slower to test a new combination of experimental geometry and reconstruction algorithm.
The approach presented at the meeting used reverse-mode automatic differentiation. Instead of manually deriving the gradient, the researcher specifies a physics-based forward model: a description of how an object produces the measured signal. Automatic differentiation then applies the chain rule in reverse through that model to calculate the gradient.
This distinction matters for technical teams because it separates two tasks: describing the forward process and calculating how changes in the model affect the objective being optimized. In the reported work, the method was demonstrated through numerical experiments in several CDI geometries:
- Transmission imaging in far-field and near-field arrangements
- Bragg coherent diffraction imaging
- Tomographic coherent diffraction imaging
The framework was implemented using established deep-learning software libraries, including TensorFlow, PyTorch, and Autograd. Its contribution was a general approach to gradient calculation across different imaging configurations, rather than a claim that one reconstruction method fits every experiment.
Machine learning in nanoscale materials modeling
The meeting also featured a CNM short course on machine learning for nanoscale materials modeling. Its stated topics included parameterizing force fields from first-principles data and determining nanoscale structure through a combination of modeling and experimental data.
These applications illustrate two different roles for machine learning. In force-field parameterization, computed data informs a model used to represent material behavior. In structure determination, experimental observations and modeling are brought together to infer nanoscale structure. In both cases, the method depends on the quality and relevance of the input data, as well as on how the modeling task is defined.
The course was intended for experimental and computational users. Hands-on participants were expected to be comfortable with a Linux command-line environment, reflecting the practical computing skills involved in this kind of analysis. Machine learning does not remove the need to understand the underlying experiment; rather, it provides tools for relating data and models in a defined research problem.
More data requires practical analysis workflows
Specialized training sessions at the meeting addressed the processing and interpretation of several kinds of X-ray data. A course on small-angle X-ray scattering covered data reduction and analysis software, including calibration, geometry, masking, plotting, fitting, and modeling. X-ray absorption spectroscopy training included data processing, corrections, principal component analysis, linear-combination fitting, and theoretical simulation.
These details highlight a practical point: advanced measurement techniques are only part of a successful workflow. Data reduction, calibration, analysis, and interpretation also require suitable tools and expertise. Different techniques need different processing steps, and the choice of method must reflect the measurement and the question being asked.
The meeting’s joint workshop on artificial intelligence, advanced data analysis, and data management in the APS Upgrade era placed these issues in the context of changing research-facility capabilities. The program identifies data analysis and management as connected concerns. As experimental programs evolve, teams need to consider not only how data is collected, but also how it will be processed, analyzed, and managed.
What technical teams can take from these methods
For engineers, system integrators, and technical decision-makers considering advanced imaging or analytics, these research examples suggest several practical questions:
- What physical process connects the measured signal to the information the team needs?
- Can that process be represented in a forward model?
- Which analysis steps are difficult to derive or repeat manually?
- What computing skills, software, and data-management practices will the workflow require?
The answers will depend on the application. The APS/CNM examples do not establish that a particular method is suitable for a specific plant or production process. They do show how researchers structure difficult analysis problems: define the measurement model, select methods that fit the data, and make processing steps explicit enough to test and refine.
Conclusion
AI-driven scientific data analysis brings together experimental knowledge, mathematical methods, and computing tools. At the 2019 APS/CNM Users Meeting, automatic differentiation offered a way to calculate gradients for phase retrieval from a specified forward model, while machine learning topics addressed force-field parameterization and nanoscale structure determination. Training in scattering and spectroscopy further underscored the importance of robust data-processing workflows. For technical teams, the practical lesson is to begin with the measurement and analysis problem—not the algorithm—and to plan for the full path from data collection to interpretation.