AI Frameworks Optimize Battery Life and EV Safety

A new hybrid review and original research article published in Discover Vehicles presents a unified framework for artificial intelligence-driven intelligent sustainable mobility. The study, authored by Hafsa Waheed and Ghulam E Mustafa Abro, integrates AI-enabled embedded architectures for electric and hybrid transport systems with original analytical insights on deep reinforcement learning (DRL) for energy management. The paper critically examines how AI-driven embedded systems collectively shape next-generation intelligent mobility ecosystems, bridging the gap between hardware, sensing, and scalable infrastructure. This comprehensive approach moves beyond isolated aspects of battery management or autonomous perception to offer a cohesive perspective on the future of electrified transportation.

The research addresses the rapid transition toward sustainable transportation, noting that existing review studies have predominantly focused on isolated components. The authors identify a lack of unified perspective on how these components interact to form intelligent mobility ecosystems. By synthesizing literature from 2009 to 2026, the paper provides a cross-domain taxonomy of AI methods across five key subsystem domains. These domains include battery technology, power electronics, battery energy storage systems, sensor fusion, and digital intelligence. The study emphasizes the need for a holistic view that accounts for the interdependencies between these subsystems in modern vehicle platforms.

A primary original contribution of the manuscript is the Thermally-Aware Twin-Delayed Deep Deterministic Policy Gradient (TA-TD3) framework for battery energy storage system (BESS) dispatch. This framework is designed to manage the thermal-economic trade-off in arid-climate deployments where ambient temperatures can reach 45 degrees Celsius or higher. The TA-TD3 architecture integrates a high-fidelity physics simulation with Arrhenius-calibrated capacity degradation and a multi-objective reward functional. It employs an adaptive thermal penalty weight that scales linearly with real-time core temperature, implementing a form of curriculum learning within the reward signal. This mechanism progressively constrains dispatch as core temperature rises, a capability that classical rule-based or model predictive control (MPC) controllers cannot replicate without explicit real-time temperature forecasting.

Evaluation of the TA-TD3 framework on multi-seasonal test episodes from the Arabian Peninsula demonstrated significant performance improvements over classical controllers. The study reports that TA-TD3 achieved a 26.6% revenue improvement, a 98.5% service life increase, and a 38.3% reduction in the levelized cost of energy relative to rule-based control. The Pareto frontier analysis showed that TA-TD3 was the sole Pareto-dominant controller, achieving the highest projected service life of 13.5 years and the second-highest annual revenue of $521.4k. These results highlight the potential of model-free policy learning to resolve complex thermal and economic constraints in grid-scale battery storage, particularly in extreme climatic conditions.

Another novel contribution is the SoC-Tracking Range-Enhancement Attachment (STRETCH), a dynamic current-limiting device for personal electric mobility (PEM) devices such as electric scooters. The STRETCH device inserts a microcontroller between the user throttle and motor controller to dynamically reduce current demand based on measured state of charge. By defining three modes of operation—full-power, moderate-limiting, and conservative—the device aims to mitigate range anxiety, a significant psychological barrier to EV adoption. The optimal breakpoint pairing, which transitions to 80% power at 25% state of charge and 40% at 5%, yielded a range improvement of 2.90% over the unmodified baseline.

The review also provides a detailed analysis of AI-enhanced battery management systems (BMS) and state estimation. It highlights the use of Long Short-Term Memory (LSTM) and transformer-based sequence models for state-of-charge estimation, which demonstrate errors below 1% under dynamic load profiles. In contrast, classical Coulomb counting methods incur errors of 5–8%. The study notes that AI models complement physics-based Arrhenius models by capturing non-linear fade acceleration in the second decade of cell life. Furthermore, the integration of digital twins (DTs) allows for predictive maintenance and adaptive control policies. These virtual models are updated via sensor telemetry and use AI synchronization layers to forecast component degradation from electromagnetic, thermal, and mechanical perspectives.

In the domain of autonomous perception, the paper analyzes AI-based sensor fusion for 3D object recognition. It discusses the integration of LiDAR, radar, and camera sensors, noting that multi-modal perception provides safer navigation than any single modality. The study identifies temporal asynchrony and spatial coordinate inconsistencies as fundamental challenges in multi-sensor fusion. For instance, a 100 ms temporal offset between sensors at highway speeds can correspond to a 3 m positional discrepancy for a moving object. The authors emphasize the need for explicit temporal alignment layers and calibration-uncertainty-aware attention mechanisms in fusion architectures to ensure safety and accuracy in real-world driving conditions.

The role of generative AI and large language models (LLMs) in mobility systems is also examined. The paper describes an LLM-based mobility stack operating across vehicle hardware, onboard edge AI, and cloud services. LLMs are being integrated for natural-language route planning, energy management, and fleet diagnostics. However, the authors identify significant risks, including hallucination, where factually incorrect but linguistically plausible outputs could have catastrophic consequences in safety-critical environments. Interpretability remains a challenge, as LLM outputs do not expose reasoning chains suitable for regulatory certification. Consequently, the study suggests that LLM deployment should initially be restricted to non-safety-critical advisory functions such as route guidance and maintenance scheduling.

Despite the advancements highlighted, the paper identifies several critical challenges and limitations. The reproducibility of AI models for state-of-charge estimation remains a concern, as most benchmarks are trained on limited chemistries and temperature ranges. Performance on these datasets may not carry over to field conditions where temperature variation and cell-to-cell variability can cause accuracy losses of 3–7 percentage points. Additionally, catastrophic forgetting in static models trained at the beginning of life is largely unsolved, and there is a lack of rigorous long-horizon validation for realistic duty cycles. The study also notes that most DRL results are based on simulated environments, and the transfer from simulation to reality remains an outstanding question requiring validation on hardware-in-the-loop platforms.

The authors conclude that the combination of advanced battery chemistry, AI-optimised wide bandgap power electronics, and multi-modal AI perception provides the technical foundation for safer and more efficient electrified transport. Future research directions include sim-to-real transfer for DRL-based BESS dispatch, federated learning for privacy-preserving fleet-wide model training, and the development of explainable AI (XAI) frameworks. These frameworks are essential to make learned AI behaviors transparent by design, addressing the black-box character of deep neural networks that currently hinders deployment in safety-critical systems. The study underscores the need for rigorous validation, standardised evaluation benchmarks, and inclusive design principles to realize the full potential of intelligent sustainable mobility.

Subject of Research: AI-driven embedded sensing, digital twins, and modular electrified vehicle systems for intelligent sustainable mobility

Article Title: Artificial intelligence-driven embedded sensing, digital twin, and modular electrified vehicle systems for intelligent sustainable mobility

Article References: Waheed, H., & Abro, G. E. M. (2026). Artificial intelligence-driven embedded sensing, digital twin, and modular electrified vehicle systems for intelligent sustainable mobility. Discover Vehicles, 2(1), Article 20. https://doi.org/10.1007/s44465-026-00025-0

Image Credits: AI Generated

DOI: 10.1007/s44465-026-00025-0

Keywords: artificial intelligence, electric vehicles, battery energy storage, deep reinforcement learning, digital twins, sensor fusion, autonomous driving, sustainable mobility, battery management, personal electric mobility, Artificial, intelligence-driven