Inverse Intelligence Group

We turn forward models upside down

If you cannot clearly assign the cause to measured data, then you are dealing with inverse problems. Our team of experts in physics and data science can help: Using invertible neural networks and scientific AI, we turn your measurement noise into understandable parameters and simulations into predictive sensors.

We make ambiguous measurement and process data from industrial plants finally analyzable - through the development and integration of AI solutions. Our open VIPR Inversion Engine delivers Bayesian inference/inversion, uncertainty quantification, virtual sensing, real-time alerting, and fully integrable REST interfaces – production-ready, transparent, and modularly extensible. We combine physical understanding, scientific machine learning, and productive implementation to help industrial companies and applied research institutions gain robust insights and control logic from the most complex data.

Inverse problems we have solved

VIPR - Versatile Framework for Inverse Problems

The Problem: Reconstruction of hidden material properties from complex measurement data

Working with the neutron reflectometer creates a physically demanding inverse problem: from noisy reflectivity measurement data, central material properties such as layer thickness, roughness, or density profiles must be reconstructed. Together with MLZ Instrument Scientist Dr. Alexandros Koutsiumpas, the entire measurement and analysis process was examined. The challenge is to read the machine's data almost in real-time, automatically select the appropriate pre-trained model from metadata, and reliably feed the analysis results back into the experimental workflow – for example, for in-situ machine control. At the same time, the results should be visualized in an understandable and user-friendly way via a frontend.

The Solution Approach: Invertible neural networks for fast and robust reconstructions

An invertible neural network enables direct solution of the inverse problem: it learns the mapping between measurement data and material properties in a way that this relationship remains unique and reversible. This allows the model to directly infer the underlying material parameters from noisy reflectivity curves – without time-consuming iterative simulations. Through automated selection of a suitable, pre-trained model, new measurement data can be analyzed immediately, while the web interface allows even non-specialists to intuitively experiment with different models and compare their results.

The Result: Dramatic acceleration and better usability in the research process

This approach achieved a 195-fold acceleration: analysis time dropped from around 185 seconds to about 1 second – fast enough for real-time applications. The results can be immediately fed into automated workflows, allowing machines to be dynamically adjusted during measurement. At the same time, the intuitive web visualization ensures that both experts and non-specialists can efficiently evaluate the data and immediately assess measurement quality. This creates a powerful and practical solution that demonstrates how modern invertible neural networks can solve complex technical challenges in practice.

Confi-GRIP - Design, Simulation and Optimization of Gripping Processes for Flexible Assembly and Order Picking

The Problem: Complex components require flexible but time-intensive gripping strategies

In industrial production, robotic grippers must regularly be adapted to new, complex, and often unknown components. Today, this task is usually done manually: specialists define a reliable, force-controlled gripping position through time-consuming trial-and-error – while simultaneously considering the correct positioning of components from KLTs, pallets, or blisters. This process is not only slow but also error-prone and dependent on individual experience. The goal was therefore to develop an AI-based software platform that flexibly designs gripper technologies, tests them virtually, and optimizes them for real production processes. A web-based configurator should automatically derive an appropriate gripping strategy from CAD data, select suitable grippers, and make necessary geometric adjustments to the design.

The Solution: AI solves inverse problems in gripper development

Flow diagram of the automated development of grippers: On the left, a CAD model of a component that is intended to be manipulated by a virtual gripper. This setup is tested in an AI simulation. An optimization loop connects the virtual gripper and the AI simulation. On the right, the final result is shown: a real, physical gripper mounted on a robot arm.

Two inverse problems are at the center: First, the optimal gripper design must be reconstructed from a component's geometry. Second, the necessary gripper parameters should be predicted from individual camera shots ("n=1"). A central web-based AI platform was developed that generates training data, trains models, and can interact directly with industrial robots. Using simulation data and CAD models, large amounts of synthetic training data are created, with which the AI precisely learns the relationship between object geometry, gripping point, force distribution, and gripper design. This is combined with AI-based path planning, seamlessly embedded in classical robot programming. Thus, the system can predict both optimal gripping strategies and the associated robot movements.

The Result: Faster to optimal gripper designs and robust production processes

The solution enables gripper technologies to be developed, adapted, and optimized in a highly automated manner rather than manually. Companies receive a tool that efficiently analyzes complex components, suggests suitable gripper geometries, and safely simulates the gripping process. Through AI-supported generation and use of synthetic training data, new components can be integrated without lengthy test series – even when only a single camera shot is available. The combination of intelligent gripping strategy, precision simulation, and automated robot path planning leads to more robust production processes, shortened development times, and significantly increased flexibility in automation.

HZwo - Digital Twin of a PEM Fuel Cell System

The Problem: Making hidden operating states in fuel cells visible

Schematic top view of a hydrogen-powered car: On the left is a hydrogen tank with H₂ inlet and outlet, in the center a fuel cell that converts hydrogen into electricity, and on the right an electric motor that drives the vehicle. All components are connected by conduits.

At Fraunhofer IWU in Chemnitz, the focus was on developing novel operating strategies for energy-efficient drives. Especially with PEM fuel cells, the internal state of the cell plays a crucial role in its lifetime and operational safety. The challenge is to draw conclusions about internal operating conditions such as temperatures or pressures from externally measurable sensor data – such as currents, voltages, or flow quantities. At the same time, anomalies must be detected early before they lead to damage or failures. For this purpose, a data platform was created that centrally collects, links, and makes sensor and simulation data usable for condition monitoring. The central question was: Which internal states can be reliably reconstructed from available data and with what accuracy?

The Solution: AI-supported reconstruction of internal states through invertible models

The underlying inverse problem consists of reconstructing the non-directly measurable operating conditions inside the fuel cell from observable sensor data. Invertible neural networks are particularly well-suited for this: they learn the reciprocal relationship between internal states and sensor data in such a way that this relationship is unique and usable in both directions. This allows the models to precisely derive how temperature, pressure, or humidity distributions look inside the cell: and immediately recognize deviations. This AI-based condition monitoring makes it possible to identify anomalies such as pressure loss at the anode early and protect the fuel cell from consequential damage.

The Result: From measured sensor data such as cell voltage and power output, defects and aging effects in the fuel cell can be detected

AI-supported analysis significantly improves the economic efficiency and operational safety of fuel cell systems: operating states can be specifically optimized, aging reduced, and failures prevented. The developed condition monitoring system is generalizable and suitable not only for PEM fuel cell stacks but also for applications in electrolysis and hydrogen mobility. This creates a scalable solution that provides clear insights into complex electrical and chemical processes – and enables companies to directly benefit from AI-based condition monitoring for their own technologies.

Bayesian Inference & INNs

Development, training, and adaptation of invertible models for layer thicknesses, material parameters, plasma states, sensor drifts, and other ambiguous measurement processes.

Data Strategy & Data Acquisition

Planning and construction of data pipelines, data models, data collection, robustification, and automated data management.

Synthetic Data & Simulation

Generation of training data using physical models or generative models – ideal when data is scarce or expensive.

VIPR Integration

Implementation of a modular analysis and control pipeline based on our open-source VIPR framework. With REST API, plug-ins, alerting, and dashboards.

Product Integration

Seamless integration into existing systems, measurement device software, automation systems, or edge infrastructure.

Interested?

Our Working Method

Problem Definition and Domain Analysis
- Formulation of the inverse or data-driven problem in physical context
- Analysis of relevant mechanisms, boundary conditions, and measurement chains
- Identification of reconstructable quantities and necessary diagnostics
Data Acquisition and Preliminary Analysis
- Structured recording of measurement, process, and metadata
- Synchronization, filtering, calibration, and quality checking
- Extraction of physically significant features and time regimes
Design of Inversion Model and Training
- Selection of suitable model architecture (e.g., Invertible Neural Network, Physics-informed Neural Network, Bayesian Inference)
- Integration of physical constraints and process knowledge
- Training with real, synthetic, or hybrid datasets, possibly on high-performance computers
System Integration and API Deployment
- Provision of the model as a scalable API or edge component
- Connection to measurement systems, controls, and data services
- Secure operation via REST, gRPC, or OPC-UA
Validation and Pilot Operation
- Experimental verification of reconstructions and predictions
- Analysis of uncertainties, robustness, and real-time capability
- Iterative adjustment based on pilot data and expert feedback
Production Deployment and Lifecycle Support
- Deployment in target environment with monitoring and versioning
- Continuous model monitoring, drift detection, and re-training
- Technical support, documentation, and continuous optimization

We are more than just another AI agency

We are not just an AI agency on cloud nine. We are physicists, machine learning engineers, scientists, innovators, and we work in a practical manner. We combine science-driven AI with a deep understanding of processes and systems. With open frameworks like our VIPR Inversion Stack, Bayesian inference, physics-informed neural networks, and normalizing flows, we build transparent, auditable, and industry-ready pipelines that seamlessly link measurement data, simulations, and automation systems. This creates solutions that enable robust state reconstructions, secure process control, and scalable integrations, far beyond what traditional AI agencies deliver.

What Distinguishes Us

Scientific Machine Learning

We develop science-driven AI models that explicitly consider physical laws, nonlinearities, and uncertainties. Our focus is on physics-informed neural networks, normalizing flows (INN/cINN), and hybrid models that combine data and process knowledge. This allows us to derive robust state reconstructions and process predictions from limited or noisy measurement data.

Bayesian Inference

For inverse problems, we use Bayesian inversion methods and invertible neural networks to determine latent states from observable effects – including uncertainty quantification. This approach allows proper treatment of ambiguous measurement situations, incorporation of prior knowledge, and statistically validated reconstruction of solution spaces. Our methods are applied where classical optimization fails or physical models are incomplete.

Understanding Complex Systems

Our team of physicists and computer scientists analyzes systems with many coupled mechanisms, nonlinear feedbacks, and emergent behavior. Whether plasma, coating, or measurement processes: we model the underlying structures in latent state space and make their dynamics quantifiable. This allows us to visualize stability limits, transition regimes, drift phenomena, and process hysteresis – and make them usable for data-driven optimization.

Open Source Frameworks

Our tools are based on open, transparent frameworks such as PyTorch or VIPR, deployed in large research facilities like MLZ. We develop modular pipelines for data acquisition, analysis, inversion, and visualization – fully traceable and extensible. Open source creates trust, facilitates audits, and enables long-term maintainability in industrial applications.

Measurement and Automation Technology

We integrate AI models into industrial measurement and automation systems. This includes:

Multi-sensor fusion
Real-time state reconstruction
REST and edge deployments
Feedback to control systems

We connect modern AI with proven measurement technology and automation to make processes more robust, adaptive, and predictable.