Traditional cyber-physical systems operate in heavily constrained and controlled environments with limited exposure to unexpected changes and uncertainties. Examples include robots operating on manufacturing assembling-lines and cyber-physical control systems of chemical plants. The model-based design paradigm, where design, implementation and verification are all guided by mathematical models of the system, has proven to be very successful in building such non-adaptive cyberphysical systems and proving their safety. The recent success of data-driven approaches based on the collection of a large amount of data followed by learning and inference has enabled modern cyberphysical systems to be more adaptive. Examples include self-driving cars and warehouse robots. Learning algorithms embedded in these systems allow them to learn as they execute and modify their behavior as needed. Such systems are capable of a wide range of non-preprogrammed behaviors. But this creates a new challenge. Model-based design paradigm is no longer sufficient. Formal guarantees on safety, robustness or improvement in performance are difficult to establish since the system evolution is no longer static; instead, it is data-driven and guided by the system's dynamic experience. The goal of this project is to build and evaluate a formal framework that combines data-driven and model-based development of adaptive cyber-physical systems. This project develops a new approach for designing safe, data-driven, and model-based adaptive cyber-physical systems (CPS). Model-based techniques are used initially to bootstrap the system and find the most liberal safety envelope for the system. A combination of design robustness and runtime monitoring of quantitatively-interpreted rich temporal logic is used to keep the system within the safety envelope. Data-driven techniques are used to actively explore, adapt, and improve system performance while constraining the system behavior to lie within the safety envelope. New data is summarized by tight learning of temporal logic properties from it; the learned logical specification is, in turn, used to guide active exploration. The key advances in this project include (a) data as model paradigm, where data from past runs is treated as a first-class object in the design of CPS, (b) tight learning from positive-only examples, where previous runs (that are all safe runs, and hence provide only positive examples) are summarized into rich temporal logic formulae, (c) safety envelope synthesis for robustness-metric guided monitoring and optimization of system performance within the envelope, (d) data-driven extensions of model-based control, where data is used to extend classical model-predictive control, and (e) active exploration, where an adaptive CPS actively executes some safe manoeuvres solely for the purpose of improving its knowledge and performance.