Organizations: Department of Information and Communication Engineering, School of Informatics, Xiamen University, Xiamen 361005, China · Key Laboratory of Underwater Acoustic Communication and Marine Information Technology (Xiamen University), Ministry of Education, Xiamen 361005, China · Navigation College, Jimei University, Xiamen, Fujian 361021, China
Accurate future channel state information (CSI) is essential for underwater acoustic (UWA) communication because delays associated with signal propagation, feedback, scheduling, and transceiver reconfiguration can render currently available CSI outdated. However, most existing predictors estimate only the next channel state and obtain multi-step forecasts through recursive rollout, which causes prediction errors to accumulate as the multipath structure evolves. The scarcity of sea-trial measurements further limits reliable model training. To address these challenges, we propose LLM4UACP, a dual-domain, structure-preserving framework that adapts a pretrained Transformer for few-shot multi-step UWA channel prediction. LLM4UACP extracts complementary features from the channel impulse response (CIR) and channel frequency response (CFR), aligns them through Fourier-consistent fusion, and encodes the fused representation as delay-time tokens. A pretrained Transformer with a frozen backbone is adapted through Butterfly Orthogonal Fine-tuning to jointly predict multiple future CIRs in a single forward pass, thereby avoiding the error accumulation associated with recursive prediction. Experiments on the measured SO-06 and SO-07 channels show that LLM4UACP achieves normalized mean squared error (NMSE) values of 0.0411 and 0.0399, corresponding to reductions of 88.53% and 85.15%, respectively, relative to the most recent baseline. With only 0.11M trainable parameters, LLM4UACP maintains robust performance across varying prediction horizons and observation-window lengths, limited-data settings, and cross-dataset transfer scenarios.
Figures & tables
Fig. 1: Sea-trial measurement geometry in the Taiwan Strait.
Fig. 2: Experimental setup of the underwater acoustic channel sounding sea trial, including the DAC-based transmission chain, ADC-based reception chain, and baseband signal processing procedure.
Parameter
Value
Transmission range
30km
Water depth
50m
Transducer depth below surface
30m
Wind speed
5 – 11m/s
Carrier frequency
3.5kHz
Symbol rate
1000bps
TABLE I: Experimental parameters of the underwater acoustic channel sounding sea trial.
Fig. 3: Measured time-delay-domain CIRs for SO-06 and SO-07.
Fig. 4: Temporal correlation of the measured CIRs for SO-06 and SO-07.
Fig. 5: Measured time-frequency-domain CFRs for SO-06 and SO-07.
Fig. 6: Frequency correlation of the measured CFRs for SO-06 and SO-07.
Fig. 7: Architecture of the proposed LLM4UACP.
Hyperparameter
Value
Tobs,Tpred
16,4
Delay taps / patch size, K , (Pt,Pd)
124 , (4,8)
Hidden dimension / layers / heads
512,6,8
BOFT block size / butterfly stages
8,4
Optimizer / batch size, B
AdamW, 128
Learning rate / weight decay, (η,λ)
(2×10−4,1×10−4)
TABLE II: Hyperparameter configuration of LLM4UACP.
Model
NMSE
Model Complexity
Inference Efficiency
SO-06
SO-07
Total Params (M)
Trainable Params (M)
FLOPs (G)
Avg. Latency (ms)
Std. Latency (ms)
RLS
0.3375
0.2989
-
-
-
0.786
0.206
AttLSTMPreNet
0.6065
0.3765
3.13
3.13
0.102
1.130
0.288
CsiPreNet
0.6397
0.4868
11.03
11.03
1.649
2.068
0.395
STL-Transformer
0.3586
0.2688
0.15
0.15
0.041
1.706
0.507
LLM4UACP
0.0411
0.0399
21.08
0.11
12.103
7.602
0.097
TABLE III: Overall Evaluation Results of LLM4UACP and Baseline Models
Fig. 8: Delay-time CIR prediction results on the SO-06 dataset.
Fig. 9: BER comparison in SO-06 (left) and SO-07 (right). The predicted channels are used for signal detection, while the true channel serves as the Perfect CSI benchmark.
Fig. 10: NMSE under prediction horizon Tpred with fixed Tobs=16 on SO-06 (left) and SO-07 (right).
Fig. 11: NMSE under observation window length Tobs with fixed Tpred=4 on SO-06 (left) and SO-07 (right).
Fig. 12: Few-shot NMSE performance under different training sample ratios on SO-06 (left) and SO-07 (right).
Fig. 13: Cross-dataset generalization. In-domain denotes same-dataset training and testing, and cross-domain denotes direct target-dataset evaluation without target-domain training. Left: SO-06 to SO-07. Right: SO-07 to SO-06.
Ablation Group
w/o Ablation Variant
Model Scale
Prediction Error
Total Params (M)
Trainable Params (M)
NMSE (SO-06)
Δ (SO-06)
NMSE (SO-07)
Δ (SO-07)
Full Model
LLM4UACP
21.08
0.11
0.0411
–
0.0399
–
Dual-Domain Feature Extraction
w/o Frequency Branch
21.07
0.10
0.0442
+7.6%
0.0412
+3.3%
w/o Delay Branch
21.07
0.10
0.0505
+22.3%
0.0474
+18.9%
w/o Linear Bypass
21.08
0.11
0.0581
+41.4%
0.0577
+44.7%
w/o Adaptive Fusion
21.08
0.11
0.0427
+3.9%
0.0413
+3.6%
TABLE IV: Grouped Ablation Study of LLM4UACP under High-Resolution Prediction Settings. Δ denotes the relative NMSE increase compared with the full model.
School of Electronic and Information Engineering, Beihang University, Beijing 100191, China · Department of Strategic and Advanced Interdisciplinary Research, Peng Cheng Laboratory · Hangzhou International Innovation Institute, Beihang University, Hangzhou +8
Ericsson Research, Stockholm, Sweden · Department of Computer and Information Science, Linköping University, Linköping, Sweden · Department of Electrical Engineering, Linköping University, Linköping, Sweden
Intelligent Networks, German Research Center for Artificial Intelligence (DFKI GmbH), Kaiserslautern, Germany · Institute for Wireless Communication and Navigation, RPTU Kaiserslautern-Landau, Kaiserslautern, Germany