Why Battery Intelligence Matters

In April 2019, a battery storage container in Surprise, Arizona, highlighted the limitations of conventional battery monitoring. Although the fire suppression system activated successfully, gas venting had stopped and external sensors indicated that the event was under control. Three hours later, when a firefighter opened the container, it exploded, seriously injuring several first responders.

Forensic analysis traced the incident to a single 18650-class cell. Lithium had plated on its anode, creating a metallic path across the separator and causing an internal short. The cell entered thermal runaway, triggering a cascade to neighbouring cells and filling the sealed container with flammable gases. Yet the monitoring system reported a battery that, according to its thresholds, appeared normal.

The suppression system itself had not failed; it was designed to extinguish a fire and it succeeded. However, thermal runaway propagation is an exothermic cascade that can generate its own oxygen while releasing hydrogen, methane and carbon monoxide. The visible fire was suppressed while the underlying cascade continued

The Limitations of Conventional Battery Monitoring

A battery monitoring system can only detect failure modes its underlying model can recognise. In this case, the sensors and thresholds were functioning correctly; the limitation was the model itself.

The battery is the only component that ages and can conceal that aging.

A battery can continue delivering performance while degrading invisibly and non-linearly. Voltage, temperature and state of charge may appear normal while internal changes such as SEI growth, lithium consumption and silicon cracking are already occurring.

This is where battery intelligence becomes critical. It goes beyond monitoring current conditions to understand what is happening inside cells, how it is changing and what it means for future performance, safety and cost. This supports battery lifecycle management, degradation prediction and predictive battery maintenance.

How Battery Management Systems Have Evolved Since 2019

The battery industry has evolved significantly since 2019. LFP chemistry accounted for more than 55% of global EV battery deployment and over 90% of stationary storage in 2025, while pack prices declined by 8%.

Battery safety standards have advanced beyond basic voltage and temperature thresholds. Connected systems now generate extensive telemetry and support over-the-air updates, cloud dashboards and remote control.

The challenge is converting this data into mechanism-specific, predictive decisions, a growing focus for modern battery management system (BMS) software and battery intelligence.

Case Study 1: The Limits of BMS Software

In late 2025, a major automaker recalled more than 320,000 plug-in hybrid vehicles following 19 battery-pack fires, including nine in vehicles that had already received an earlier software remedy. Regulators concluded that the previous diagnostic system could not detect certain battery abnormalities involving separator damage capable of causing an internal short.

The vehicles were connected, monitored and supported by diagnostic software, yet the system could not adequately identify the underlying failure mechanism. The missing ingredient was not more data, but a model grounded in cell physics. This is where battery modeling becomes essential, providing insights beyond what conventional battery management systems (BMS) can infer from surface-level measurements.

Case Study 2: Lithium Plating and Battery Safety

In 2026, a compact electric vehicle sold across several continents was recalled after a cell-manufacturing deviation was associated with lithium-plating growth and potential internal shorts. As an interim precaution, owners were instructed to limit charging to 70% and avoid parking vehicles indoors while affected modules were inspected and replaced.

Despite modern technology, newer-generation cells and connected systems, the failure mechanism remained electrochemical. Lithium plating is influenced by cell-to-cell variation, temperature, current, state of charge and operating history.

A fleet-wide charging restriction can reduce immediate risk, but cannot identify which individual cells or packs are approaching a critical condition. For lithium battery applications, including systems supported by a lithium-ion battery management system, battery intelligence can distinguish healthy from deteriorating cells and enable more targeted, actionable interventions.

Case Study 3: From Remote Mitigation to Predictive Battery Maintenance

In late 2025, certain residential battery-storage systems were remotely discharged after potentially defective cells were identified. Approximately 10,500 units in one market were involved following 22 overheating events, including five fires. Remote discharge helped contain the immediate hazard while replacement arrangements were made.

However, remote discharge is a mitigation strategy implemented after the affected population has been identified. It is not equivalent to detecting the physical signature of an emerging defect directly from operational battery data.

The next step is understanding how a failure develops and prescribing the safest intervention. For an energy storage system or battery energy storage system (BESS), this means moving beyond alerts toward battery optimization and actionable intervention.

Why Battery Data Alone Is Not Enough

There is an increasingly common assumption that batteries are primarily a data problem: collect enough telemetry, feed it into a sufficiently large neural network, and meaningful insights will emerge. The reality is more complex. Machine learning is highly effective at interpolation, while batteries frequently require extrapolation.

A model trained on three years of fleet data may not have experienced the conditions of year eight, a new chemistry, a different fast-charging profile or temperatures outside its training data. These are precisely the conditions where predictive accuracy matters most.

Battery degradation is also path-dependent. Two battery packs with identical cumulative throughput can have very different health trajectories. One may have experienced prolonged exposure to high state of charge and temperature, leading to SEI growth and lithium inventory loss, while another may have undergone aggressive charging at low temperatures, increasing the likelihood of lithium plating and loss of active material.

Another challenge is limited ground-truth data. Battery packs are rarely dismantled during normal operation to provide labelled mid-life degradation data. Without reliable ground truth, purely supervised approaches risk becoming sophisticated forms of curve fitting rather than genuine representations of battery behaviour.

Combining Battery Modeling and Battery AI

The solution is not to choose between physics and artificial intelligence. Physics provides the structure. AI provides the correction.This combination forms the foundation of advanced battery AI and digital twin technology.

An electrochemical core can use a P2D/Doyle–Fuller–Newman formulation or a reduced-order single-particle model incorporating electrolyte dynamics. These models can represent physical state variables such as:

• Lithium concentration gradients
• Overpotentials
• Anode plating margin
• SEI thickness

These physical quantities are governed by relationships such as conservation laws rather than simply statistical features, allowing them to remain meaningful when operating conditions move beyond historical datasets.

Machine learning can estimate changing parameters, learn the residual between the physical model and real-world behaviour and improve the model as new operational data becomes available. This can include physics-informed neural networks, hybrid state observers and Bayesian identification using differential-voltage and incremental-capacity signatures from field charging curves. The outcome is a battery digital twin capable of providing significantly more than a conventional dashboard.

From Battery Monitoring to Actionable Intelligence