Human-centered battery twins

BAT-GPT

Can a battery digital twin explain its condition in language a user can act on?

Published in the 2025 IEEE ICUFN proceedings

Overview

How can a parameter-efficient language model make battery digital-twin insights more interactive and understandable?

Battery twins produce technical measurements and predictions, but their value depends on whether engineers and asset owners can interpret them. BAT-GPT explores a conversational layer that translates battery operating data into direct, domain-specific responses.

The project fine-tunes FLAN-T5 with Low-Rank Adaptation using battery conversations derived from NASA data. Zero-shot, few-shot, and parameter-efficient fine-tuning experiments reveal both the promise and the current limitations of language models as digital-twin assistants.

BAT-GPT model-in-the-loop battery digital twin and language assistant architecture

The complete BAT-GPT architecture links a capacity-predicting model-in-the-loop digital twin to prompt-response generation, FLAN-T5 adaptation, real-time battery data, and an interactive response loop.

Methods

Battery conversations

Battery cycles, current, temperature, voltage, and capacity are transformed into prompt-response examples grounded in digital-twin data.

LoRA fine-tuning

FLAN-T5 is adapted to the battery domain through low-rank parameter updates rather than full-model retraining.

Prompting and evaluation

Base, zero-shot, few-shot, and fine-tuned responses are compared using human baselines and ROUGE metrics.

4ROUGE measures reported

0.298LoRA-FLAN-T5 ROUGE-L

0.292LoRA-FLAN-T5 ROUGE-Lsum

LoRAparameter-efficient adaptation

Contributions

A domain-specific conversational assistant for battery digital twins.
A prompt-response pipeline grounded in battery operating cycles and environmental conditions.
Parameter-efficient adaptation of FLAN-T5 using LoRA.
Comparison of base, zero-shot, few-shot, and fine-tuned behavior.
An honest evaluation that documents improved coherence alongside remaining numerical errors.

Battery dataset prompt generation and digital twin conversation example — Battery measurements are converted into grounded prompt-response pairs, then used to teach the assistant how to explain capacity under specific current, temperature, cycle, and operating conditions.

Interpretation

BAT-GPT is valuable because it tests the interface between numerical prediction and human understanding. The evaluation does not hide failure: some responses become clearer after adaptation while capacity estimates can remain inaccurate. That evidence defines where grounding and validation must improve.

Current scope

This study is an early language-interface prototype built around NASA battery data and FLAN-T5. It establishes the conversational component later integrated into the broader BatteryMetrix vision.

Additional figures

FLAN-T5 encoder-decoder architecture used by BAT-GPT — The FLAN-T5 encoder-decoder architecture provides the language backbone, using attention and feed-forward blocks to transform battery prompts into contextual responses.

Low-rank adaptation method used to fine-tune BAT-GPT — LoRA freezes the pretrained language-model weights and learns compact low-rank update matrices, making domain adaptation more parameter efficient.

Battery digital twin interface with conversational assistant panel — The prototype places an assistant beside the battery twin so users can query model results within the monitoring environment.

Comparison of baseline human, original FLAN-T5, and parameter-efficient model responses — A response comparison shows how the base and parameter-efficient models interpret a battery-capacity prompt against a human reference.

Zero-shot FLAN-T5 response to a battery operating-condition prompt — The zero-shot example exposes repetition and weak physical interpretation, documenting the limitations that motivate domain-specific adaptation.