AI-Based Systems: Novel Method Defines Operational Conditions from Existing Data

Researchers are increasingly deploying Artificial Intelligence in safety-critical applications, yet defining the precise operational conditions for these systems remains a significant hurdle, particularly when relying on existing data. Johann Maximilian Christensen, Elena Hoemann, and Frank Köster, from the Institute for AI Safety and Security at the German Aerospace Center (DLR), alongside Sven Hallerbach et al., address this challenge with a novel approach to defining the Operational Design Domain (ODD) , a crucial requirement for system certification. Their work introduces a ‘Safe-by-Design’ method that defines the ODD a posteriori from collected data, utilising a multi-dimensional kernel-based representation, and importantly, demonstrates the potential to equalise data-driven ODDs with the original, underlying conditions of the data itself. This research offers a pathway towards the certification of data-driven, safety-critical AI systems, validated through both simulation and a real-world aviation collision-avoidance scenario.

This research introduces a multi-dimensional, kernel-based representation to a posteriori define the ODD, a description of the environmental conditions in which an AI system must operate. Traditionally, constructing the ODD relies on expert knowledge and established standards, a process that becomes increasingly difficult for complex real-world systems or when utilising existing datasets. The team achieved a deterministic framework for data-driven ODD construction, utilising samples as anchor points for analytically defined kernel functions, offering a continuous and bounded representation.

This breakthrough reveals a method applicable even with sparse data, making it suitable for both early development stages and mature systems. Researchers formalised ODDs as mathematical structures, introducing a kernel-based affinity representation and an automated procedure for parameter selection and handling of out-of-distribution samples. By construction, the resulting ODD representation is deterministic and interpretable, enabling its use for modelling, monitoring, and future certification processes for AI-based systems.
Experiments show the approach is validated through both Monte Carlo methods and a real-world aviation use case focused on a future safety-critical collision-avoidance system. The research further defines conditions for determining the equality of two ODDs, proving that a data-driven ODD can accurately reflect the original, underlying hidden ODD present within the data itself. This capability is crucial for ensuring the reliability and safety of AI systems operating in complex and unpredictable environments. The work opens possibilities for streamlining the certification process for data-driven, safety-critical AI-based systems, reducing reliance on extensive testing and post-hoc analysis.

Moreover, the study unveils a framework that moves beyond reactive safety measures towards proactive guarantees integrated directly into the system development process. By explicitly defining operational conditions, the ODD provides a foundation for system design and verification, aligning with evolving regulatory frameworks. The novel kernel-based ODD representation captures complex parameter dependencies and naturally reflects real operational conditions, offering a dynamic and adaptable approach to safety assurance. This research addresses the challenge of defining environmental conditions for artificial intelligence systems, particularly in complex real-world scenarios where complete environmental descriptions are often unavailable. The team engineered a deterministic framework that constructs ODDs directly from data samples, employing these as anchor points for analytically defined kernel functions. This approach yields a continuous and bounded ODD representation, uniquely determined by the available data and applicable even with limited datasets.

Researchers harnessed Monte Carlo methods to validate the accuracy of this data-driven ODD construction technique. Experiments employed simulated data to assess the method’s ability to accurately represent the underlying hidden ODD from which the data originated. Furthermore, the study pioneered a real-world aviation use case, focusing on a future safety-critical collision-avoidance system, to demonstrate the practical applicability of the kernel-based ODD. Scientists implemented an automated procedure for parameter selection and handling of out-of-distribution samples, ensuring the robustness and reliability of the method. This innovative approach contrasts with traditional ODD creation, which relies on expert knowledge and can be difficult to apply to complex systems. This work formalised ODDs as mathematical structures, introducing a kernel-based affinity representation that allows for a precise and interpretable definition of operational conditions. The technique reveals complex parameter dependencies inherent in the data, naturally reflecting real operational conditions and facilitating updates as new data become available. The research team validated this approach using both Monte Carlo methods and a real-world aviation use case focused on a future safety-critical collision-avoidance system. Results demonstrate the ability to a posteriori define the ODD, a description of the environment in which an AI system operates, even when traditional expert-based methods are difficult to apply. This breakthrough delivers a data-driven approach to ODD construction, offering advantages over manual creation, particularly for complex real-world systems.

Experiments revealed that the kernel-based representation accurately captures complex parameter dependencies within the operational environment. The team measured the ODD’s ability to equal the original, underlying hidden ODD of the data, confirming the method’s fidelity in reconstructing the true operational space. Measurements confirm that the data-driven ODD is continuous, bounded, and order-independent, crucial properties for safety-critical applications. Tests prove the method’s applicability even with sparse data, making it suitable for both early development stages and mature systems with limited information.

The study formalised ODDs as mathematical structures, introducing a kernel-based affinity representation for precise definition. Scientists recorded that the automated procedure for parameter selection and handling of out-of-distribution samples enhances the robustness and reliability of the ODD construction process. Data shows the resulting ODD representation is deterministic and interpretable, enabling its use for modelling, monitoring, and future certification of AI-based systems. The work introduces a novel approach to ODD construction, moving away from reactive measures like extensive testing and towards proactive safety guarantees integrated into the system development process. This advancement enables future certification of data-driven, safety-critical AI-based systems by providing a verifiable and well-defined operational space. Traditionally, the ODD, which describes the conditions in which a system is designed to operate, is established early in development using expert knowledge. This research introduces a data-driven approach utilising a multi-dimensional kernel-based representation to define the ODD retrospectively from existing data. The method was validated using Monte Carlo simulations and a real-world aviation scenario focused on a future collision-avoidance system.

The key achievement lies in demonstrating that a data-driven ODD can, under certain conditions, equal the original, underlying ODD that generated the data. This is particularly significant for safety-critical applications where certification requires a clearly defined ODD. The kernel-based representation also allows for the creation of ‘soft’ safety boundaries, providing graded warning zones and enabling systems to degrade gracefully or alert operators before reaching critical limits. Authors acknowledge that a fully truthful reconstruction of the ODD from data is not always achievable. Future research will concentrate on optimising the selection of kernel parameters and exploring extensions to model dependencies between dimensions.

Integrating temporal information to represent evolving ODDs is also planned, potentially benefiting systems in dynamic environments like air traffic management and autonomous driving. Further work will focus on integrating this kernel-based ODD approach into formal certification workflows, linking affinity thresholds to safety requirements and enabling controlled data-driven updates. These developments aim to strengthen the use of data-driven ODDs for deploying safe and reliable AI-based systems.