Compositional Machine Learning for Time-Verifiable, Safe, and Explainable Cyber-Physical Systems

The integration of Machine Learning (ML) into safety-critical Cyber-Physical Systems (CPSs) presents a fundamental challenge: how to leverage the performance benefits of ML while ensuring the rigorous safety, timing, and explainability requirements of critical applications. Traditional monolithic machine learning approaches create blackbox systems that are di!cult to verify, implement on hardware for timing analysis, and explain to stakeholders. This thesis addresses these challenges by introducing a novel paradigm of compositional machine learning that builds ML-based CPSs that are safe, time-verifiable, and explainable by construction. In this thesis, we introduce the systematic decomposition of monolithic models into smaller models. This decomposition preserves functional behaviour while enabling parallel execution and modular verification. We develop novel compilation frameworks that translate Python-based Artificial Neural Network (ANN) models to both VHDL (for Field Programmable Gate Array (FPGA) implementation) and C code (for embedded systems), enabling Worst-Case Execution Time (WCET) analysis. Our compilers specifically support compositional ANN architectures, addressing a critical gap in existing tools. Moreover, we introduce a policy-driven framework for safe-by-construction ML systems that mines safety properties from data and uses them to guide the training of compositional models. Additionally, our runtime enforcement mechanism ensures policy adherence in real-time, providing a safety net for unexpected scenarios. Finally, we develop a novel compositional explainability paradigm combining multiple explainability techniques to provide actionable insights at sub-model and system levels. We validate our framework through comprehensive case studies in autonomous vehicles and healthcare, demonstrating generalisability, practical applicability and concrete benefits. Through extensive evaluation, we demonstrate that compositional ANN models can achieve up to 85% reduction in WCET, 53% reduction in hardware resources, and 40% reduction in computations while maintaining comparable performance to monolithic approaches. Furthermore, we illustrate how a compositional policy-mining approach helps obtain ML models that are safer than their monolithic counterparts. Finally, we also show how compositional ML models are inherently more explainable than monolithic models. Overall, the results establish that compositionality enables the development of ML-based CPSs that are not only more performant but also more verifiable, implementable, and trustworthy.

Name thesis:degree_name

PhD

Level thesis:degree_level

Doctoral

Discipline thesis:degree_discipline

Electrical and Electronic Engineering

Grantor dc:publisher

ResearchSpace@Auckland

Year dc:date.issued

2026

Author dc:creator

• Chatterjee, Sobhan

Advisors dc:contributor.advisor

• Roop, Partha
• Patel, Nitsh

• Formal methods and Software (461203)
• Machine Learning (4611)
• Safety

Statement dc:rights

• Items in ResearchSpace are protected by copyright, with all rights reserved, unless otherwise indicated.

Licence dc:rights.uri

• https://researchspace.auckland.ac.nz/docs/uoa-docs/rights.htm

Handle dc:identifier.uri

https://hdl.handle.net/2292/76158

OAI identifier oai:identifier

oai:researchspace.auckland.ac.nz:2292/76158