Artificial intelligence at Fraunhofer IIS

Machine learning, mathematical optimization, neural networks, deep learning – the methods of artificial intelligence open up brand new opportunities and represent a paradigm shift for applied research.

Why do we use artificial intelligence? Through the use of AI, we can develop solutions that were previously inconceivable. Building on our many years of experience in microelectronics, sensor technology and data analysis with numerous outstanding developments, we employ artificial intelligence in all our research areas. In this way, we develop new technologies and services to meet the pressing challenges of industry, commerce and society with smart and secure solutions.

What distinguishes our work? We combine outstanding know-how in the individual specialist subject areas with a holistic view of AI usage, covering everything from hardware and sensor technology, to data communication and data analysis, to the optimization of individual applications and entire business processes. 

A brief guide to the main AI concepts

Definitions of the most important technical terms in artificial intelligence

What is artificial intelligence? Artificial intelligence (AI) is a branch of computer science devoted to endowing machines with capabilities that resemble intelligent (human) behavior. This can be achieved by means of preprogrammed rules or by machine learning.

What is explainable AI? The decisions arrived at in black box models, especially in deep artificial neural networks, cannot be understood by humans. Explainable AI seeks ways of making the hidden logic or individual output easier to reconstruct and explain.

What is machine learning? Machine learning refers to methods whereby an algorithm or machine learns regularities from data samples, or learns through repetition how to execute a task in an increasingly better manner with reference to a defined quality criterion

What is mathematical optimization? Mathematical optimization aims to detect all optimization potential for a specific planning problem using a selected evaluation criterion, and then to choose the best solution from among many possible ones.

This can be a question of maximization or of minimization, such as the maximization of production output or the minimization of energy consumption. Mathematical optimization takes into account all constraints that need to be fulfilled so that the solution can be actually implemented (security conditions, logistical dependencies, etc.).

In terms of methodology, mathematical optimization proceeds as follows: the specific planning problem is first translated into a mathematical model, so that efficient algorithmic methods can be deployed to derive the best solution for the selected optimization goal. The results are then translated back into the language of the user, so that they can assess and implement the solution in practice.

What is neuromorphic hardware? When experts talk about neuromorphic hardware, they mean brain-inspired computers or individual components of a computer that mimic or model neurological or artificial neural networks. To do this, they use highly connected parallel synthetic neurons and synapses.

This hardware configuration is many times more adept at calculating deep neural networks (DNNs) than computers with classic von Neumann architecture.

What are neural networks? Artificial neural networks are a basis for machine learning methods modeled on nerve cell connections in the brain. They consist of data nodes and weighted connections between them. Machine learning processes make changes to various parameters in the network and are thus able to optimize it for specific tasks.

What is deep learning? Deep learning refers to neural networks with a very high number of levels that make it possible to tackle new classes of problems.

What is weak artificial intelligence? Weak artificial intelligence is able to master a very specific task particularly well, such as playing chess, recognizing texts, evaluating images and finding regularities and patterns in large data sets. Tasks that go beyond this special job exceed the ability of this form of artificial intelligence. The kind of artificial intelligence used today comes within the sphere of weak AI.

What is strong artificial intelligence? Strong AI, also known as general AI, refers to machine intelligence that possesses human-like cognitive abilities. This means it can accomplish feats such as transfer learning or intelligently combining several tasks such that they are optimally coordinated with each other. Experts disagree on whether such a strong form of artificial intelligence can ever be developed.

Institute-wide Fraunhofer IIS projects in the AI domain

New solutions through AI


ADA Lovelace Center for Analytics, Data and Applications

The ADA Lovelace Center provides a unique research infrastructure for connecting research and business in Bavaria. Methods, techniques and expertise from the fields of data analytics and AI are employed for specific practical issues and also further developed as a result of these efforts. Through close collaboration and innovative cooperation formats between industry and research, companies are given access to comprehensive AI expertise and quickly accrue concrete benefits, such as a sustainable improvement in business processes and higher quality of decision-making.

Go to the ADA Lovelace Center

Foto zum Projekt ADA-Center
© dragonstock -

Near-sensor AI

Artificial intelligence helps us evaluate large volumes of data. But is all this data really necessary? Near-sensor AI means that the sensor can already make an objective evaluation of the quality and information content of the data that it processes. For the purposes of data economy, AI directly at the sensor helps ensure that only the data that is genuinely useful gets into the system, thereby greatly increasing the quality of the data actually collected.

More about near-sensor AI

Künstliche Intelligenz - KI
© Fraunhofer IIS/Udo Rink

Neuromorphic hardware

Neuromorphic hardware refers to specialized circuits for the efficient calculation of AI algorithms. As such, it represents a highly powerful approach for the realization of AI in many distributed and energy-saving nodes, such as in the Internet of Things (IoT) and autonomous driving. Neuromorphic hardware helps not only to reduce power consumption and costs, but also to substantially reduce latency times in signal processing.

More about neuromorphic hardware

Icon image - neuromorphic hardware
© zapp2photo –

Center for Digital Signal Processing using Artificial Intelligence – DSAI

In the DSAI, Fraunhofer IIS combines digital signal processing with artificial intelligence methods. This vastly expands the capabilities of digital signal processing, making it possible to tackle previously unsolvable challenges in various application domains such as mobility, consumer electronics, intelligent assistance systems or telecommunications.

More about the DSAI


Künstliche Intelligenz - KI
© greenbutterfly - / Fraunhofer IIS

Further AI projects at Fraunhofer IIS by research area

We harness and expand the potential of artificial intelligence for a wide variety of sectors and fields of application.

Audio and Media Technologies

  • In the KISS project, we are collaborating with the Friedrich-Alexander-Universität Erlangen-Nürnberg to investigate new development tools that will allow us to improve AI-based algorithms in signal processing and translate them into highly efficient computer-assisted implementations. Through our research, we are helping to reduce development times by a significant margin and to facilitate optimization efforts and new developments of products and services for mobility, communication and entertainment. We will be sharing our research results with all relevant stakeholder groups through lectures at universities and in the form of further education and training for industry.

  • Together with Fraunhofer IAIS, we are developing a voice assistant that not only meets European standards of data security, but also elevates the quality of human-machine communication to a whole new level. Its semantic capabilities will far outstrip the abilities of existing systems from the United States or Asia. In this way, we are improving the possibilities of human-machine interaction through language and ensuring that companies can use voice assistants with full data sovereignty and in compliance with European data protection standards.

Engineering of Adaptive Systems

  • We are guiding Germany into the AI era. In addition to other leading initiatives in the sphere of AI, we are participating in AI4Germany to actively support the domestic economy and German society as a whole in the application of artificial intelligence. 

  • To prepare manufacturing companies for the requirements of the future, new concepts are needed to help them master the rapid growth of digitalization and connectivity of processes. Our principal focus here is on the automatic analysis of measurement and process data for intelligent condition monitoring and quality assurance. After all, companies are already collecting vast amounts of data today, but in many cases without extensively evaluating it. With efficient structuring, self-learning systems and novel big data and AI approaches, this dormant potential can be transformed into strategic competitive advantages.

  • Industry has long been using wireless systems in all kinds of different areas. And for pioneering applications such as cooperation between mobile robots, wireless communication is actually a key technology. Its growing importance for critical industrial processes, however, is absolutely contingent upon wireless systems having a high level of availability and reliability, something that is currently often not the case.

    Through machine learning methods, it is possible to detect and avoid looming system failures, irrespective of the complexity of systems or of the wireless technology used.

  • Capturing the data produced by machines and systems is generating ever greater quantities of data. To help customers efficiently use this data, our Engineering of Adaptive Systems (EAS) research area supports them in the development of intelligent sensors that handle the preprocessing of information on a decentralized basis directly at the machines. So as to further optimize the analytical hardware, the integration of algorithms from the domain of artificial intelligence is becoming increasingly widespread. For their design to be optimal, however, there needs to be particularly high-performance, flexible and energy-efficient electronics in place. To this end, we use novel concepts and advanced packaging methods for hardware development – all the way down to modular structures with chiplets that permit the incorporation of individual assemblies. By means of intelligent IPs, the EAS research area also offers a solution for the automation of analog neuromorphic hardware design. Customized generators enable the fast and reliable generation of building blocks for schematics and layout, which speeds up the design process considerably.

  • In Project KIKiS (Artificial intelligence – Expertise and Innovation Potential in Saxony), our Engineering of Adaptive Systems (EAS) research division in Dresden collaborated with the Technische Universität Dresden to investigate the needs and requirements for the development and deployment of AI in companies. You can download the corresponding research paper on our website. It contains recommendations for companies and various actions that can be undertaken to develop Saxony into a strong AI region.  

Development Center X-ray Technology EZRT

  • We invest in intelligent production monitoring and control through X-ray technology (2D and computed tomography). In this way, we detect deviations from the optimal production process at an early stage and regulate the process such that the specified component or product quality is always achieved, meaning that there are no rejects. Reject-free production increases resource efficiency, reduces costs and is better for the environment.  

  • We employ AI methods to analyze and evaluate X-ray images of material flows (mineral ores from mines, recycling materials, etc.). This information can then be used to make a sorting decision or calculate the value of the material flow. 

  • Through the use of three-dimensional, nondestructive monitoring systems, we are able to capture data from plants in a highly complete and precise manner without damaging them. Artificial intelligence then classifies and analyzes the various plant structures. On this basis, we can investigate plant yield and stress resistance under different climatic conditions. 

Communication Systems

  • To make autonomous driving safe, the vehicle needs utterly reliable technologies that enable it to directly and precisely capture the given traffic situation and its own position within the traffic. Together with our project partners, we are developing the high-performance, energy-efficient AI FLEX hardware platform and the corresponding software framework for autonomous driving on behalf of the German government. With the help of artificial intelligence, AI FLEX will quickly and reliably process and pool data from laser, camera and radar sensors in cars.

Positioning and Networks

  • As digitalization progresses, more and more data is being generated. We employ AI methods such as machine learning so that we can continue to draw the right conclusions from this growing mass of information. With the help of AI, we recognize regularities and patterns in the data and use them to optimize not only the sensor systems, but also entire processes. 

  • At the Research Center IoT-COMMs, we investigate and develop connected, agile and mobile production systems and applications for autonomous driving. To this end, we combine the technologies of networks, positioning and information security and use AI to realize learning, cognitive sensor technology.
    The Research Center IoT-COMMs is part of the Fraunhofer Cluster of Excellence Cognitive Internet Technologies (CCIT).

  • Higher quality and flexibility in assembly: to support manual work processes in assembly, we develop cognitive sensors that can be fitted to handheld tools as an add-on module. They record things like whether a screw was inserted in the correct place. Through machine learning, we facilitate the integration of intelligent tools into production.

  • In the “R2D – Road to Digital Production” project, we fit tools with smart production tags. This enables the tools to autonomously recognize process steps, log them and control them in coordination with the other workpieces. As a result, traditional production processes are more dynamic and more decentralized. 

  • Through smart objects, we are facilitating intelligent and transparent processes for learning production and logistics. Process mining and machine learning enable smart objects to control processes themselves, all the way down to batch sizes of one.

Smart Sensing and Electronics

  • In the domain of affective sensing / affective computing, we are working on giving machines the ability to recognize, interpret and process human emotions and affects. By refining sensor and algorithmic technologies, we are able to capture physiological responses caused by agitation, excitement and other moods in real time. This includes things like the recognition of facial expressions and the interpretation of biosignals.

  • The goal of our research is to create new methods of explainable machine learning in conjunction with prediction and prescription (forecasting and explanation) in the medical and automotive application domains. Such methods include those intended not only to ensure transparency in the training and workings of deep neural networks (DNNs), but also to improve the validation capability of forecast model content. An additional focus is on the generalizability of machine learning systems; that is, on the development of adaptive methods that can handle individual patient fluctuations in physiological parameters (e.g. mobile use while playing sports) and under varying conditions (»in-the-wild«).

  • The ADELIA project is part of a new research funding concept in Germany on the subject of energy-efficient AI systems.  

    For many applications, and especially for battery-powered mobile devices, current AI systems consume far too much power. Together with our colleagues from Fraunhofer IPMS, our focus is on designing an analog inference accelerator ASIC. Our team combines interdisciplinary experience from the fields of low-power IC development, medical engineering, machine learning and signal processing, all of which are key to developing an energy-efficient ASIC for AI applications. 

  • The ERIK project aims to develop a robotic platform as a physical interaction partner and development tool for treating children with impaired socio-emotional functioning, such as children on the autism spectrum. Our approach encompasses the integration of sensor-based emotion recognition based on facial expressions, language and physiological signals; the creation of innovative child-robot interaction using several therapeutic interaction approaches; and the boosting of autistic children’s ability to communicate through play-based training and positive reinforcement (e.g. for successful learning).

  • The goal of the overall KI-PREDICT project, which is sponsored by the German Federal Ministry of Education and Research (BMBF), is to apply the methods of artificial intelligence (abbreviated KI in German) to various levels of the production process. This forms the basis for the condition-based predictive maintenance of production facilities and the monitoring of product quality directly in the manufacturing workflow.
    Our focus within this project is on developing a sensor interface ASIC. What makes this technology special is that is designed specifically for condition monitoring and real-time process control, and facilitates energy-efficient feature extraction and signal processing directly at the sensor.

  • A key element of psychotherapy for anxiety disorders is to expose patients to situations that trigger the anxiety. The Optapeb project aims to develop a system that records the emotional responses of clients in a multimodal manner during such exposure and employs data fusion to extract parameters that are relevant for the further course of treatment. Micro-interventions are then derived from these parameters and provided to the client by a virtual agent in an intuitive interaction. Through machine processing of the data sets acquired in numerous exposures, the system arrives at suggestions for successful micro-interventions.

  • The aim of the TraMeExCo project is to research and develop suitable new methods for robust and explainable AI (XAI) in complementary fields of application (digital pathology, pain analysis, cardiology) in the domain of medical engineering. To this end, we investigate and implement approaches such as few-shot learning and heat maps for digital pathology and determine heart rate variability in noisy ECG and PPG data using long short-term memory networks. Based on the Facial Action Coding System (FACS) as developed by Ekman and Friesen, the researchers investigate Bayesian deep learning methods using pain videos. Meanwhile, our project partner the University of Bamberg focuses on the conception, implementation and testing of methods aimed at the explainability of diagnostic system decisions using LIME, LRP and ILP methods.

  • Explainable AI is viewed by many as the key topic for AI research going forward, and explainability is a necessary condition for the deployment of machine learning in practice. In the Comprehensible AI (CAI) project group, we are collaborating with the University of Bamberg to develop methods for explainable machine learning, as purely data-driven machine learning cannot be employed in many application fields or else only in combination with other methods. In this way, we are developing methods of interactive and incremental learning, working on hybrid approaches and developing algorithms for the generation of multimodal explanations.

Supply Chain Services

  • We are carrying out research aimed at continuously improving AI analytics methods and techniques further, so that they can be used to obtain manageable and qualitative data and information from seemingly unmanageable data volumes. The goal is to extract the right data, the information with concrete commercial benefits.

  • Many decision-makers are currently faced with questions concerning how they should digitalize their company, and how they must change their business model and/or organizational structure so as to remain commercially viable and successful in the future.

    We help companies find the right answers to these questions. Enterprises need fresh perspectives on their business so that they can continually reinvent themselves in a fast-changing digital world. They obtain these perspectives on their processes, organizational structures and business models by integrating people – whether customers, users or employees – into the process of analysis, while developing data-driven, forward-looking and market-oriented services.

  • Increasing connectivity calls for new concepts for data spaces and technology deployment. As data generally comes from different sources and systems and often exists in a wide variety of formats, it needs to be processed before it can be actually leveraged as information to bring about optimization in a networked environment. This requires methods and technologies suitable for the preparation and integration of data. To respond to this need, we are developing new methods and techniques by means of which data can be modeled and prepared such that it is easy to use in new contexts; namely, by removing the data from its existing environment while retaining the same depth of detail.