Reinforcement learning squeezes earthquake-detecting AI onto tiny edge devices

Deep learning models that pick out seismic phases—the arrivals of P-waves and S-waves in ground motion recordings—demand substantial computation, which makes them difficult to run on small, low-power hardware. A study published in Earth Science Informatics on 26 October 2026 proposes a method called TLDQ, short for Time-Layer Joint-Aware Reinforcement Learning Dynamic Quantization, which adaptively decides how precisely each part of a seismic phase-picking network should be represented as it processes a signal. Quantization, the practice of storing a model's numerical weights with fewer bits than the full floating-point standard, is a practical way to reduce storage and computational costs.

The problem, the researchers note, is that fixed-bit-width quantization treats every part of the network and every moment of the signal identically. Seismic waveform complexity varies from one temporal segment to the next, and different layers of a network exhibit different quantization sensitivities. A uniform bit width therefore risks compromising picking accuracy.

TLDQ tackles this by making bit-width selection a sequential decision problem solved with Q-learning, a form of reinforcement learning in which an agent learns a policy through repeated trial and reward. At each step, the agent observes three things: the complexity level of the current time segment of the seismic signal, the sensitivity level of the network layer it is working on, and how much of the computational budget remains. Guided by reward feedback, it iteratively refines its bit-width allocation over a candidate set of bit widths.

The team applied the method to two seismic phase-picking models, EQTransformer and SeisLM. Average bit widths fell to 4.12 bits for EQTransformer and 3.34 bits for SeisLM. Despite this aggressive trimming, performance held up: for both models, P-wave accuracy and F1-score remained above 97 percent, and the S-wave F1-score stayed above 93 percent.

The efficiency gains were substantial. Model storage size dropped by up to 89.6 percent, and BitOps—a measure of the computational work involved in inference—fell by up to 87.4 percent. To test the approach under real conditions, the researchers deployed the quantized EQTransformer on a Raspberry Pi 5B, a compact edge device.

There, the quantized model achieved 94.54 percent accuracy, only 0.87 percentage points below the full-precision original, while cutting the average inference time to 89.66 milliseconds—a 5.3-fold speedup. The system also ran stably for 72 consecutive hours, indicating that the approach can sustain continuous operation rather than just passing a benchmark. What makes the approach conceptually notable is the joint awareness it builds in.

Because the agent weighs temporal signal complexity and layer-wise sensitivity together with the remaining budget, it can allocate bits adaptively: complex waveform segments and sensitive layers can retain more precision, while simpler segments and robust layers are compressed harder. The result is a bit-width allocation tailored to both the input signal and the network architecture, rather than imposed uniformly. The study's scope is specific: two phase-picking models, quantized and evaluated on benchmark tasks and on a single edge device.

The reported results are measured outcomes of those experiments, not demonstrations across other model architectures or hardware platforms. Within that scope, the authors present TLDQ as an effective solution for accurate, efficient and reliable seismic phase picking in resource-limited environments. With a demonstrated 5.3-fold speedup, sustained 72-hour stability on commodity edge hardware and accuracy losses measured in fractions of a percentage point, the approach shows that dynamic, reinforcement-learning-guided quantization can bring deep learning-based phase picking within reach of resource-constrained devices.

Subject of Research: Reinforcement learning-based dynamic quantization for seismic phase picking on resource-constrained edge devices

Article Title: TLDQ: time-layer joint-aware reinforcement learning dynamic quantization algorithm for seismic phase picking on resource-constrained edge devices

Article References: Cai, J., Zhou, J., Shi, Y., & Cui, D. (2026). TLDQ: time-layer joint-aware reinforcement learning dynamic quantization algorithm for seismic phase picking on resource-constrained edge devices. Earth Science Informatics, 19(12), Article 228. https://doi.org/10.1007/s12145-026-02281-w

Image Credits: AI Generated

DOI: 10.1007/s12145-026-02281-w

Keywords: seismic phase picking, dynamic quantization, reinforcement learning, Q-learning, edge computing, EQTransformer, SeisLM, Raspberry Pi, earthquake detection, model compression, BitOps, deep learning