Hybrid LSTM-GRU-CNN Emulator

• The LSTM-GRU-CNN emulator is a hybrid neural model combining LSTM for long-term dependencies, GRU for efficient short-term memory, and CNN for local feature extraction.
• It emulates computationally intensive processes such as 21-cm brightness temperature reconstruction, marine biogeochemistry forecasting, and speech recognition.
• The design leverages stacked meta-learning, parallel branching, and sequential compositions to drastically reduce computation time while maintaining high predictive accuracy.

An LSTM-GRU-CNN emulator is a hybrid neural architecture that combines Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and Convolutional Neural Network (CNN) components to model structured signals with both sequential and local or spatial organization. In the most direct sense, the term denotes a surrogate for computationally expensive simulation, as in reconstruction of X-ray effects on the global 21-cm brightness temperature during the Epoch of Reionization, where a stacked hybrid model couples base predictors with an LSTM-CNN-GRU meta-model (Hosseini et al., 7 Aug 2025). Closely related hybrids also appear as emulators, forecasters, and end-to-end predictors in marine biogeochemistry, automatic speech recognition, missile trajectory prediction, and environmental time-series modeling, indicating that the same architectural vocabulary is used across scientific inference, signal processing, and sequence prediction (Skakala et al., 25 Jun 2026, Dhakal et al., 2024, Baradaran, 20 Mar 2025, Tavares et al., 2024).

1. Terminological scope and problem classes

In the cited literature, “emulator” is used most explicitly for models that approximate a simulator or process-based numerical system. The 21-cm reconstruction model is trained on the 21SSD simulation and is intended to reconstruct the effect of X-ray flux on the global 21-cm brightness temperature without repeated simulations; it is reported to preserve temporal and spatial dependencies, generalize to unseen parameter combinations, and reduce computation time by a factor of one million while preserving excellent prediction accuracy of 99.93\% and an error margin of less than 0.35 mK (Hosseini et al., 7 Aug 2025). In marine biogeochemistry, deep-learning emulators are used to reproduce a high-complexity pelagic model within a one-dimensional water-column framework, with LSTM and physics-informed 1D CNN variants evaluated for daily forecasting and multi-decadal stability (Skakala et al., 25 Jun 2026).

A broader, looser usage is visible in tasks that are not simulator surrogates in the strict sense but still instantiate the same hybrid principle. These include Nepali automatic speech recognition with 1D-CNN, ResNet, and bidirectional LSTM combinations (Dhakal et al., 2024); hypersonic missile trajectory prediction with parallel CNN, LSTM, and GRU branches (Baradaran, 20 Mar 2025); monthly fire-spot forecasting in the Amazon using a sequential LSTM-GRU model that the paper explicitly notes can be generalized by adding CNN layers for multivariate or spatial-temporal data (Tavares et al., 2024); speech emotion recognition through an ensemble of 1D-CNN, LSTM, and GRU models (Ahmed et al., 2021); Vietnamese hate-speech detection via a Bi-GRU-LSTM-CNN classifier (Huynh et al., 2019); gene mutation classification with a parallel LSTM-BiLSTM-CNN-GRU ensemble (Aburass et al., 2023); and image captioning with CNN-GRU encoder-decoder reconstruction (Ahmad et al., 2023).

This distribution of applications is consistent with the comparative overview of CNN, LSTM, and GRU models: CNNs are described as strong for spatial dependencies, whereas LSTM and GRU are described as better suited to sequential and time-series structure; hybrid CNN-RNN models are therefore motivated where both local pattern extraction and temporal context are material to the target mapping (Shiri et al., 2023).

2. Architectural organizations

The architectural family is not a single template but a set of recurrent-convolutional compositions. One recurring pattern is stacked meta-learning. In the 21-cm emulator, a base ensemble composed of a Deep Neural Network and Random Forests produces preliminary predictions and residuals, and a stacked hybrid meta-model then refines them with LSTM, CNN, and GRU components. The reported configuration uses LSTM with 512 units and L2 regularization, CNN with 288 units and L1 sparsity followed by max pooling, and GRU with 512 units and L2 regularization, together with dropout and early stopping (Hosseini et al., 7 Aug 2025).

A second pattern is parallel branching with late fusion. The hypersonic trajectory model feeds the same input sequence into three branches: a 1D convolutional branch with 64 filters and kernel size 3, an LSTM branch with 64 units, and a GRU branch with 64 units. The flattened branch outputs are concatenated, followed by a dense layer with 128 neurons and a linear output; three separate models are trained for the $x$, $y$, and $z$ coordinates (Baradaran, 20 Mar 2025). The gene mutation classifier follows a comparable parallel design in text space: GloVe embeddings feed LSTM, BiLSTM, CNN, and GRU branches, whose outputs are aggregated before a dense softmax classifier (Aburass et al., 2023). The Vietnamese hate-speech model also uses a hybrid branching design after the embedding layer, combining CNN-1D, Bi-LSTM, and Bi-GRU modules before classification (Huynh et al., 2019).

A third pattern is sequential recurrent composition. The Amazon fire-forecasting model uses a fixed input window of 12 months, followed by an LSTM layer with 256 units, a GRU layer with 256 units, a dense layer with 256 units and ReLU activation, and a final linear output (Tavares et al., 2024). In speech recognition, the best Nepali ASR system uses a different sequential arrangement: five stacked residual blocks based on 1D-CNN, batch normalization, PReLU, and skip connections, followed by two bidirectional LSTM layers, dense layers, and a CTC output head (Dhakal et al., 2024).

A fourth pattern is ensemble specialization rather than a monolithic hybrid. In speech emotion recognition, three models are trained separately—1D CNN + FCN, 1D CNN + LSTM + FCN, and 1D CNN + GRU + FCN—and their softmax outputs are combined by weighted averaging (Ahmed et al., 2021). This is distinct from the parallel-branch formulation, but it expresses the same division of labor between local feature acquisition and long-range dependency modeling.

Several papers make the role decomposition explicit. CNNs are used for local feature extraction, localized spatial features, or spatial pattern extraction; LSTMs are used for long-range temporal dependencies; and GRUs are used either as a computationally efficient sequence model or as a complementary recurrent mechanism emphasizing short- to medium-term dependencies (Hosseini et al., 7 Aug 2025, Baradaran, 20 Mar 2025). In residualized CNN-RNN systems, the skip pathway is written as

$\text{Output} = G(x) + x,$

which is used to mitigate degradation as depth increases (Dhakal et al., 2024).

3. Input representation and preprocessing

The input side of LSTM-GRU-CNN emulation is domain dependent, and much of the reported performance is tied to careful preprocessing. In Nepali ASR, the model is trained on the OpenSLR Nepali corpus with 157,905 audio clips from 527 speakers sampled at 16 kHz. Silent gaps at both ends of most clips are clipped by a custom sliding-window algorithm, a step reported to reduce dataset length by more than half. Audio is represented with Mel Frequency Cepstral Coefficients using 13 mel scales per frame and hop size 40, yielding a 52-dimensional feature vector per time frame; the mel-scale transform is written as

$$m = 2595 \cdot \log_{10}\left(1 + \frac{f}{700}\right).$$

The output vocabulary contains 66 tokens, including Nepali characters and padding, unknown, and blank tokens (Dhakal et al., 2024).

In the 21-cm emulator, the data come from the public 21SSD database. The simulations span a $200~h^{-1}~\mathrm{cMpc}$ volume with $1024^3$ particles and vary three parameters: X-ray Efficiency $f_X \in \{0.1, 0.3, 1, 3, 10\}$, Hard-to-Soft X-ray Ratio $r_{H/S} \in \{0, 0.5, 1\}$, and Lyman Band Emissivity Efficiency $f_\alpha \in \{0.5, 1, 2\}$, over 400 redshift bins from $y$0 to $y$1. After preprocessing, approximately 18,000 datapoints remain. A distinctive design choice is residual feature engineering: residuals between base-model predictions and true values are introduced as new features after being found highly correlated with the target but uncorrelated with the original inputs; all features are normalized to zero mean and unit variance (Hosseini et al., 7 Aug 2025).

In the Amazon fire model, the input is a univariate monthly series of active fire spots from June 1998 to August 2024, with windows of length 12 used to predict the next month. The split preserves temporal order: June 1998 to August 2021 for training, September 2021 to August 2023 for validation, and September 2023 to August 2024 for testing (Tavares et al., 2024). In hypersonic trajectory prediction, each sample is a multivariate time series of position coordinates $y$2 generated by numerical integration of equations accounting for gravity, drag, lift, and atmospheric density, and the model predicts the next position from a sliding input window (Baradaran, 20 Mar 2025).

Speech emotion recognition uses a handcrafted feature vector of dimension 155 composed of 128 Log-Mel Spectrogram coefficients, 13 MFCC coefficients, 12 Chromagram coefficients, 1 Zero-Crossing Rate coefficient, and 1 Root Mean Square coefficient. Data augmentation is performed before feature extraction by Additive White Gaussian Noise, pitch shifting, and stretching the signal level (Ahmed et al., 2021). In NLP and biomedical text tasks, the input representations are embedded token sequences: Vietnamese hate-speech detection uses FastText pretrained Vietnamese vectors with sentence length capped at 220 tokens (Huynh et al., 2019), while gene mutation classification uses tokenized clinical text padded to length 512 and mapped through pre-trained GloVe embeddings (Aburass et al., 2023). These examples indicate that the “CNN-LSTM-GRU” label does not determine the representation level; the same hybrid family operates on audio features, state variables, tabular simulation outputs, trajectory coordinates, and embedded text.

4. Objectives, decoding, and regularization

The optimization criterion varies with the mapping being emulated. For unaligned sequence transduction, the Nepali ASR model uses Connectionist Temporal Classification (CTC) for loss calculation and CTC beam search decoding for character prediction, allowing training without frame-level alignment between speech and text (Dhakal et al., 2024). For regression-oriented emulation, Mean Squared Error is standard; in the 21-cm model it is the optimization target,

$y$3

with Mean Absolute Error and $y$4 also reported (Hosseini et al., 7 Aug 2025). The hypersonic trajectory model likewise uses MSE between predicted and actual positions, optimized with Adam (Baradaran, 20 Mar 2025).

Forecasting models in the time-series literature also use error-based regression criteria. The Amazon fire model reports MAE and RMSE for selection and evaluation, trains for 1000 epochs with Adam, and uses He normal initialization (Tavares et al., 2024). In marine biogeochemistry, the reduced-order LSTM emulator is trained with MAE, whereas the physics-informed 1D CNN emulator predicts state increments through

$y$5

and minimizes a total loss that combines prediction MAE, mass-conservation penalties for carbon, nitrogen, and phosphorus, and bound-enforcement terms for physical limits and non-negativity (Skakala et al., 25 Jun 2026).

Classification-oriented hybrids use standard probabilistic objectives. The Vietnamese hate-speech model is trained with cross-entropy loss (Huynh et al., 2019). The gene mutation classifier uses sparse categorical crossentropy together with accuracy, precision, recall, F1-score, and Mean Squared Error (Aburass et al., 2023). The speech emotion recognition ensemble uses categorical cross-entropy, and its final prediction is the weighted average of model outputs:

$y$6

Hyperparameters and ensemble weights are tuned by grid search (Ahmed et al., 2021).

Regularization strategies are correspondingly diverse but recurrent. The 21-cm emulator uses dropout, batch normalization, L1/L2 penalties, early stopping after 10 epochs with no validation improvement, and Bayesian optimization for hyperparameter tuning (Hosseini et al., 7 Aug 2025). The hypersonic model applies dropout of 0.3 in each branch and after the dense layer (Baradaran, 20 Mar 2025). The Nepali ASR system uses dropout of 25\% in BiLSTM layers (Dhakal et al., 2024). The marine LSTM emulator uses early stopping with patience 20 and learning-rate reduction on plateau, while the marine CNN uses gradient clipping and autoregressive 10-day rollouts during training to improve stability (Skakala et al., 25 Jun 2026).

5. Reported performance and comparative behavior

Reported outcomes show that hybridization can be highly effective, but not uniformly. In the 21-cm emulator, the reported test metrics are MSE $y$7, MAE $y$8, $y$9 on test and $z$0 on train, with all result deviations from simulation below $z$1; traditional simulation time of approximately $z$2 CPU hours is contrasted with an emulator runtime of 15 minutes on consumer hardware, yielding a speedup factor of $z$3 (Hosseini et al., 7 Aug 2025). In marine biogeochemistry, the LSTM emulator is described as very stable over 22-year simulations and highly skilful for 10-day forecasts, outperforming persistence for most variables by days 2–4 at latest and achieving RMSE improvement of 50–60\% compared to persistence at day 10, whereas the physics-informed 1D CNN is generally stable but more sensitive to noise and can perform worse than persistence for most variables in short-term forecasting (Skakala et al., 25 Jun 2026).

In Nepali ASR, empirical comparison is especially instructive. A BiLSTM baseline achieves a test CER of 19.71\%, 1D-CNN + BiLSTM yields 24.6\%, 1D-CNN + ResNet + BiGRU yields 29.6\%, 1D-CNN + ResNet + LSTM yields 30.27\%, and the best model—1D-CNN + ResNet + BiLSTM with 1.55M parameters—achieves 17.06\% CER (Dhakal et al., 2024). The result shows that simply adding CNN layers does not guarantee improvement, but adding residual blocks to the CNN-BiLSTM pipeline can produce the best reported configuration for that dataset.

Other task domains show similar dependence on composition. The hypersonic missile CNN-LSTM-GRU model reports RMSE 3330.46, MAE 3110.06, and MAPE 1.83\%, with predicted and true trajectories described as closely aligned (Baradaran, 20 Mar 2025). The Amazon LSTM-GRU model reports, for the two tested seeds, RMSE Test values of 10140 and 5992 and MAE Test values of 5991 and 3785; when retrained on all data, the MAE is 2893 for seed 2024 and 2488 for seed 2025, and the model is reported to capture seasonality well, with average MAE around 2500, about 28\% of average monthly fires (Tavares et al., 2024). In speech emotion recognition, the weighted ensemble reports weighted average accuracies of 99.46\% on TESS, 95.42\% on EMO-DB, 95.62\% on RAVDESS, 93.22\% on SAVEE, and 90.47\% on CREMA-D (Ahmed et al., 2021).

Text and biomedical applications also report measurable gains from hybridization. The Bi-GRU-LSTM-CNN hate-speech classifier achieves an F1-score of 70.576\% on the VLSP public test set, compared with 69.293\% for Bi-GRU-CNN and 56.512\% for TextCNN (Huynh et al., 2019). The gene mutation classifier reports train accuracy 0.806, validation accuracy 0.615, train/validation precision 0.816/0.619, train/validation recall 0.806/0.615, train/validation F1 0.831/0.600, training time 267 seconds, and validation MSE 2.60; the same paper reports weaker performance for BERT and for an Ensemble Roberta + GloVe + LSTM baseline (Aburass et al., 2023). In image captioning, CNN+GRU with semantic reconstruction is reported to outperform the LSTM-A5 baseline across BLEU@1–4, METEOR, ROUGE-L, and CIDEr-D, for example with BLEU@4 of 0.335 versus 0.325 and CIDEr-D of 1.035 versus 1.002 (Ahmad et al., 2023).

6. Interpretive themes, misconceptions, and reported directions

One recurring misconception is that deeper or more heterogeneous hybridization is automatically beneficial. The Nepali ASR results do not support that view: BiLSTM alone outperforms 1D-CNN + BiLSTM, and only the addition of residual blocks yields the best reported CER. This suggests that depth requires explicit stabilization mechanisms rather than naive stacking (Dhakal et al., 2024). A related point appears in the 21-cm emulator, where the strongest performance is obtained not from a single hybrid stack alone but from a staged design in which base-model predictions and residuals are used as features for a higher-capacity meta-model (Hosseini et al., 7 Aug 2025).

A second misconception is that GRU should always dominate because it has fewer parameters. The comparative survey reports that GRU slightly exceeds LSTM on IMDB and is marginally best on ARAS while training faster, which supports its efficiency advantage (Shiri et al., 2023). However, the Nepali ASR experiments report worse results for BiGRU-based and unidirectional LSTM-based hybrids than for BiLSTM-based hybrids, indicating that the optimal recurrent unit remains task dependent (Dhakal et al., 2024). The literature therefore supports a narrower claim: GRU is often a practical choice for sequential modeling, but not a uniformly superior substitute for LSTM.

A third theme concerns the meaning of “spatial” within these hybrids. In astrophysical and marine emulation, CNN layers are used to extract localized spatial features or vertical-profile structure, while LSTM and GRU capture redshift evolution or temporal memory (Hosseini et al., 7 Aug 2025, Skakala et al., 25 Jun 2026). In text and audio, the same CNN component acts on token or feature sequences to extract local motifs, n-gram-like patterns, or short-range acoustic structure (Huynh et al., 2019, Ahmed et al., 2021). This suggests that the hybrid family is better understood as a mechanism for joining local receptive-field processing to gated sequence memory than as a fixed solution tied to any one modality.

Reported future directions remain domain specific. For the 21-cm emulator, the paper identifies expanding training data to underrepresented parameter regimes, incorporating hard physical priors or domain-specific constraints into the loss function, and generalizing the approach to other cosmological or astrophysical emulation tasks (Hosseini et al., 7 Aug 2025). For marine biogeochemistry, the reported prospect is extension to three-dimensional regional applications and eventual coupling with ocean physics emulators, with explainability methods used to identify key drivers of emulator behaviour (Skakala et al., 25 Jun 2026). For environmental time series, the Amazon fire paper explicitly notes that adding a CNN layer before LSTM and GRU is a useful extension for multivariate or spatial-temporal data (Tavares et al., 2024). Collectively, these directions indicate that the LSTM-GRU-CNN emulator is not a settled architecture but a configurable design space spanning stacked, parallel, residual, physics-informed, and ensemble formulations.