Search Results (315)

This paper presents a digitally controlled current-mode synaptic weighting cell in 28 nm FD-SOI CMOS for low-complexity Multi-Layer Perceptron (MLP) inference in event-based Non-Intrusive Load Monitoring (NILM). A compact bias-free [16,16] MLP is trained offline by backpropagation for fixed-weight feedforward inference. Using 616 ON/OFF events extracted from high-frequency REDD measurements, six appliance classes are characterized by four event-level features: active-power variation, reactive-power variation, current total harmonic distortion of the differential event signature, and event interval. With 16-level input quantization, the model achieves 92.9–93.4% test accuracy and 90.5–90.7% test Macro-F1, requiring 320 weighted-sum branches across two hidden layers. A four-input first-hidden-layer weighted-sum unit is selected as a representative circuit instance. Its computation is mapped to bounded current ranges using current-coded inputs, an 8-bit magnitude-controlled current-mode multiplier, sign-bit current steering, differential current accumulation, and signed-current scaling. The circuit contribution focuses on the weighted-sum datapath, particularly the repeated synaptic weighting cell; activation and complete classifier implementation are outside the scope of this work. Transistor-level PVT and supply-variation simulations validate the signed weighted-sum path, while post-layout extraction evaluates the repeated multiplier cell. The results demonstrate the block-level feasibility of digitally programmable current-mode synaptic weighting for compact event-based NILM inference. Full article

Despite an increasing interest in the pharmacology of cannabinoids from Cannabis sativa, little is known to date about their ability to cross the placenta and to be secreted into breast milk, and in this study, we sought to fill this gap. In total, 126 phytocannabinoids previously detected in Cannabis sativa were investigated for their transplacental transfer and secretion into breast milk. Placental transport was predicted using novel multiple linear regression (MLR), artificial neural network (ANN), boosted trees (BT), and support vector regression (SVR) models, based on a reference set of 84 compounds for which the placental clearance index (CI) relative to antipyrine is known. Secretion into breast milk was predicted using newly developed classification models based on soft independent modeling of class analogies (SIMCA) and One-Class Partial Least Squares (OC-PLS) algorithms. Analysis of the Q vs. Hotelling’s T² plot for the cannabinoids indicated that they are similar in their physicochemical properties to compounds empirically demonstrated to enter breast milk (“in-class”); only 7 of 126 compounds were borderline (with elevated Q but not T²); no compounds were classified as “out-of-class”. The mean predicted CI values for phytocannabinoids investigated in this study ranged from 0.4 to 0.85. It was concluded that all the cannabinoids in the studied group might cross the placenta (although their passage might be expected to be more difficult than that of antipyrine) and enter breast milk. These results should support informed risk assessment and prioritization of cannabinoids for future experimental testing. Full article

Background: Endometriosis and lipedema are chronic female-predominant disorders characterized by persistent pain that is frequently disproportionate to anatomical lesion burden. Although traditionally interpreted within distinct lesion-centered frameworks, both conditions exhibit striking clinical and epidemiological parallels, including hormonally modulated symptom dynamics, overlap with central pain syndromes, weak correlation between structural disease severity and pain intensity, and symptom clustering during reproductive transitions such as puberty, pregnancy, and menopause. Methods: This study aims to synthesize clinical, molecular, neuroimmune, and endocrine evidence on the interrelationship between endometriosis and lipedema, and to propose a hypothesis-generating neuroimmune framework linking both conditions. This integrative narrative review conducted a non-systematic literature search in PubMed/MEDLINE, Scopus, and Web of Science, focusing on mechanisms related to chronic pain, mast cell biology, TRPV1 signaling, CGRP-mediated neurogenic inflammation, intracrine steroidogenesis, and peripheral and central sensitization. Results: The review identifies convergent biological characteristics between the two diseases, including mast cell activation, macrophage polarization, endothelial dysfunction, fibrosis, angiogenesis, intracrine estrogen metabolism, and persistent inflammatory signaling. In endometriosis, direct evidence demonstrates increased sensory innervation, nerve growth factor expression, TRPV1 sensitization, CGRP-positive fibers, and mast cell-nerve interactions. In lipedema, convergent upstream mechanisms, including mast cell infiltration, elevated histamine levels, adipose tissue inflammation, and local estrogen activation, support the plausibility of a functionally analogous neuroimmune organization, despite incomplete direct neural characterization. In this context, the mast cell-TRPV1-CGRP axis is proposed as a biologically plausible framework, directly supported in endometriosis and currently hypothetical in lipedema, connecting peripheral sensitization, neurogenic inflammation, hormonal chronodependence, and central nociceptive amplification. The model further conceptualizes pain crises as transient events of instability within a sensitized neuroimmune network and proposes mechanistic phenotypes that integrate gastrointestinal, inflammatory, central, and hormonal triggers. Conclusion: Endometriosis and lipedema may represent topographically distinct manifestations of a shared neuroimmune process operating within hormone-sensitive tissues. Although the evidentiary basis remains asymmetric, with stronger mechanistic support in endometriosis than in lipedema, this framework provides a biologically plausible and experimentally testable model integrating endocrine, immune, neural, and vascular contributors to chronic pain amplification. This perspective supports coordinated translational investigation across reproductive biology, endocrinology, and pain medicine and may contribute to future mechanism-based stratification and therapeutic development. This work is hypothesis-generating and is not intended to establish causality or to provide clinical recommendations; all proposed mechanistic and therapeutic inferences require prospective experimental validation. Full article

Sleep-related respiratory disturbances are difficult to monitor continuously outside specialized laboratories because conventional polysomnography is resource-intensive and intrusive. This study presents a contactless edge-AI engineering prototype for detecting controlled voluntary respiratory-motion suppression and motion artifacts using a 60 GHz frequency-modulated continuous-wave radar. The system integrates a 60 GHz radar front end, lightweight local preprocessing, an INT8 one-dimensional convolutional neural network deployed on the Analog Devices MAX78000 CNN accelerator (Analog Devices Thailand, Chon Buri, Thailand), and an event-driven Raspberry Pi Zero 2W gateway for alert transmission. Evaluation was performed using a controlled healthy-volunteer dataset consisting of normal breathing, voluntary breath-holding-induced respiratory suppression, and deliberate motion artifact. The final valid test set contained 270 technically valid 30 s windows balanced across the three classes. The INT8 model achieved an overall accuracy of 92.6% (95% confidence interval: 88.8–95.2%), with a macro-averaged precision, recall, and F1-score of 92.6%, 92.6%, and 92.5%, respectively. Active CNN inference on the MAX78000 consumed 0.152 ± 0.011 mJ and was completed in 5.20 ± 0.11 ms, corresponding to approximately 280-fold lower active inference energy than Python 3.14.6/TensorFlow Lite 2.21.0-based execution on the Raspberry Pi Zero 2W. These results demonstrate the feasibility of privacy-aware, low-power respiratory-pattern classification at the edge. However, the study should be interpreted strictly as an engineering proof-of-concept based on controlled voluntary breathing and movement tasks in healthy volunteers. It is not a clinically validated apnea or obstructive sleep apnea detection system and did not include polysomnography, oxygen saturation measurement, airflow sensing, sleep staging, or diagnosed patient cohorts. Full article

Analog in-memory computing (AIMC) has emerged as a promising approach to mitigate the Von Neumann bottleneck in matrix operations, which are common in deep learning applications. However, the practical implementation of resistive crossbar arrays is limited by challenges in signed weight representation, conductance quantization, and device nonlinearity. This paper presents a differential mixed-signal architecture for accurate signed matrix–vector multiplication (MVM), integrated with a RISC-V microcontroller for edge inference applications. A structured digital-to-analog mapping framework encodes quantized neural network weights into programmable conductance values while preserving arithmetic correctness. The design employs voltage-mode input encoding, differential current summation, and transimpedance-based readout followed by analog-to-digital conversion, enabling single-cycle signed accumulation without duplicating crossbar resources. A 32 × 16 dual-layer prototype crossbar was fabricated and experimentally characterized. Measurements demonstrate a mean absolute percentage error (MAPE) below 1% within the linear operating region and below 4% over the full-scale conductance range. These results validate the robustness of the proposed mapping methodology and confirm the feasibility of hybrid analog–digital acceleration for edge AI systems. Consequently, this discrete prototype serves as a physical verification platform for the AIMC approach, providing valuable insights for more efficient mixed-signal computing integrated circuit (IC) designs. Full article

Memristors are state-of-the-art electronic elements with nano sizes, about 3 nm dimensions, with very good nano-second switching and memory properties, low power usage of about 100 µW, and good compatibility with the current technology of CMOS-integrated chips and circuits. These components are potentially applicable in T-byte memory arrays, artificial neural networks, logic gates and many other digital and analog electronic schemes and devices for artificial intelligence. This paper presents the application of some simple and fast-operating modified memristor models with activation thresholds in neural networks and logic circuits. MATLAB ver. 2016a and LTSPICE ver. XVII products are used for the analysis of memristor neural nets and logical circuits for artificial intelligence. Several simple, accurate and fast-operating existing modified memristor models, together with several frequently used standard memristor models, are utilized for the associated analyses and simulations. A comparison between the used memristor models is conducted. The considered memristor models are tuned, using experimentally recorded i-v relations of tungsten-sulfide Knowm memristors. An accurate functioning of the analyzed neural nets and logic functions is confirmed by the derived results. The considered modified memristor models, neural networks and logic schemes are important in modeling and analysis of memristor-based circuits for ultra-high-density artificial intelligence-integrated chips. Full article

The convergence of artificial intelligence (AI), explainable AI (XAI), and neuroscience is fostering new opportunities for understanding both machine and biological intelligence through interpretable and human-centered learning paradigms. In this Perspective, we introduce XAI2Brain as a conceptual framework for brain–AI alignment, positioning mechanistic interpretability as an intermediate layer connecting neural network representations, human understanding, and neuroscience-inspired AI design. Rather than viewing XAI solely as a post hoc transparency tool, we emphasize its emerging role in enabling mechanistic analysis of internal model representations, concept-level reasoning, and interactive human–AI alignment. We define XAI2Brain as a multi-level conceptual framework rather than a deployable system, explicitly aimed at structuring brain–AI alignment across representation-level, mechanism-level, and interaction-level perspectives. We survey the evolution of XAI methodologies—from feature attribution and concept-based explanations to mechanistic and human-centric interpretability approaches—and discuss how these methods may support bidirectional knowledge transfer between AI systems and cognitive neuroscience. Importantly, we adopt a cautious stance on brain–AI analogy, explicitly recognizing that artificial neural representations are not equivalent to biological neural representations, and instead focusing on functional and informational correspondences rather than structural equivalence. Unlike conventional human-in-the-loop or reinforcement learning from human feedback paradigms that primarily optimize behavioral outputs, XAI2Brain focuses on cognitively interpretable and mechanistically grounded alignment between AI systems and human reasoning processes. This alignment promotes interactive human-in-the-loop intelligence, empowering humans to comprehend, guide, and refine AI systems, while enabling AI systems to better interpret human instructions, intentions, and contextual reasoning. We further discuss the challenges of scaling explainability to large generative and multimodal models, including issues of interpretability robustness, cognitive compatibility, evaluation, and ethical accountability. We also highlight key limitations of current mechanistic interpretability methods, including explanation instability, representation superposition, and lack of causal guarantees, underscoring that these challenges remain open research problems. Rather than proposing a complete artificial brain architecture, this Perspective outlines a research roadmap toward more interpretable, adaptive, and neuroscience-inspired AI systems capable of supporting future brain–AI integration and collaborative intelligence. We additionally clarify that this work follows a narrative perspective review methodology with structured thematic synthesis of the literature. By framing explainability as a bridge between mechanistic AI understanding, cognitive science, and human-centered interaction, XAI2Brain highlights the importance of interpretable alignment for the next generation of brain-inspired AI systems. Full article

Convolutional neural networks (CNN) have transformed visual recognition, yet robust geometric reasoning, reliable out-of-distribution generalization, and recognition from limited data remain substantially unsolved. CNNs draw their architectural inspiration from the mammalian visual cortex, but the translation from biology to engineering was selective and, in places, imprecise, and those imprecisions have consequences that are well documented. This paper examines where the biological fidelity holds and where it gives way, grounding the analysis in formal results that predate deep learning and in recent empirical findings on CNN failure modes. We identify three diagnosable architectural limitations. First, CNNs conflate visual modalities that the biological system separates structurally at the lateral geniculate nucleus, feeding raw RGB pixels into a single undifferentiated filter bank and entangling orientation, color, and texture signals from the first layer onward. Second, CNNs repeat a spatial subsampling operation across the full depth of the network, far beyond the early visual cortex stages where it has biological warrant. Barnard and Casasent established formally in 1990 that this operation discards positional information irreversibly at every layer where it is applied, and repeating it into regions that correspond to V4 and inferotemporal cortex compounds this loss without the compensating transition to qualitatively different computations that the biological hierarchy performs. Third, the pooling-as-complex-cell analogy that motivated this design reflects a misreading of what complex cells compute. The spatiotemporal energy model formalizes complex cell behavior as geometry extraction: detecting the presence and orientation of a local edge structure robustly, abstracting over photometric accidents of contrast polarity and sub-wavelength phase that are not geometrically meaningful. Pooling is a tolerable first-stage approximation of this behavior, but as a general-purpose invariance mechanism repeated across the full depth of the network, it is attempting something categorically different, namely object-level position invariance through spatial subsampling, which achieves its goal by discarding exactly the geometric information that the energy model preserves. Treating pooling as a scalable, indefinitely repeatable implementation of complex cell behavior—rather than as a first-stage approximation with a natural biological endpoint at V3—conflates two operations that differ not in degree but in kind, and crucially it removed the principled criterion for confining the S-C operation to early visual cortex: because pooling was understood as a general-purpose invariance mechanism, the field had no architectural reason to stop repeating it. We survey how capsule networks, group-equivariant CNNs, PDE-based networks, and vision transformers each address one or two of these limitations while leaving the others intact. We propose six desiderata that a more biologically complete architecture would need to satisfy and argue that satisfying them requires treating the visual cortex’s solution as a coherent package in which each component depends on the others working correctly, rather than as a menu of independently selectable principles. Full article

The rapid growth of artificial intelligence (AI) workloads is reshaping semiconductor design across architecture, interconnect, memory hierarchy, packaging, timing, and design automation. Rather than converging on a single hardware solution, the field is expanding into a heterogeneous ecosystem that includes data-center graphics processing units (GPUs), edge neural processing units (NPUs), and application-specific integrated circuits (ASICs), field-programmable gate array (FPGA)-based and hybrid AI system-on-chip (SoC) platforms, chiplet-enabled systems, and emerging beyond-conventional-silicon approaches such as photonic, neuromorphic, and analog in-memory processors. This paper presents a comprehensive review of AI-on-chip systems from a cross-layer perspective. It examines AI chip architectures and hardware platforms, network-on-chip (NoC) designs for AI communication patterns, and algorithm–hardware co-design methods for model acceleration, including compression, quantization, and sparsity-aware optimization. It also reviews clocking, synchronization, and clock-domain-crossing (CDC) challenges in large heterogeneous systems and chiplets, as well as manufacturing, advanced packaging, and reliability issues, including two-and-a-half-dimensional (2.5D) and three-dimensional (3D) integration, thermal and mechanical constraints, assembly quality, and long-term yield considerations. In parallel, the paper surveys the growing role of AI in chip design itself, covering machine-learning-assisted analysis, Bayesian and reinforcement-learning-based optimization, and the emerging use of large language models (LLMs) and AI agents for register-transfer level (RTL) generation, design-space exploration, and autonomous electronic design automation (EDA) workflows. Finally, it discusses beyond-silicon AI chip directions and the broader economic and industry context shaping cloud, on-premises, and edge deployment. By integrating these topics into a unified framework, this review highlights the key technological drivers, system-level tradeoffs, and future research directions that will define next-generation scalable, reliable, and energy-efficient AI-on-chip systems. Full article

Processing-in-memory and in-memory computing (PIM/IMC) are increasingly explored to mitigate the von Neumann data-movement bottleneck that limits deep neural network (DNN) performance and energy efficiency. Progress, however, remains fragmented across device substrates, architectural prototypes, mapping and scheduling methods, compiler toolchains, and benchmarking practices, making results hard to compare and slowing deployment. This survey synthesizes developments from 2019–2025 along four coupled axes: (i) memory substrates and architectural design, (ii) mapping, partitioning, and scheduling, including learning- and graph-based strategies, (iii) compilers and end-to-end deployment flows, and (iv) benchmarking datasets, metrics, and reporting norms. Drawing on over twenty representative platforms spanning static random-access memory (SRAM) and dynamic random-access memory (DRAM), emerging non-volatile, capacitive, and photonic substrates, we clarify the trade-offs separating analog/charge-domain IMC from digital SRAM/DRAM-centric PIM, including reported peaks up to 600 TOPS/W and 1.5 TOPS/mm². We organize mapping frameworks into a unified reference taxonomy, identify recurrent evaluation pitfalls that undermine reproducibility, and highlight persistent gaps in training support, robustness under non-idealities, and coverage of large-scale GNN workloads. Finally, we outline a five-phase roadmap from benchmark standardization to industrial validation toward compiler-integrated, GNN-informed PIM/IMC systems validated on production-scale workloads. Full article

As software systems increasingly combine machine learning, deep learning, and generative AI components with classical deterministic logic, the systematic detection of AI-based algorithmic elements in application code is becoming essential for software audit, compliance with the EU AI Act (Regulation (EU) 2024/1689), and quality assurance. This paper introduces Quality-Informed Neural Networks (QINN), an architecture in which the structured knowledge encoded in the Quality Function Deployment (QFD) House of Quality is embedded into the network topology and weight initialisation through QFD-derived binary structural masks and knowledge-calibrated initialisation—in direct analogy with Physics-Informed Neural Networks (PINNs). The QFD relationship matrices act as structural priors that constrain the hypothesis space toward quality-consistent solutions by enforcing domain-expert-validated sparsity on network connectivity, while an optional QFD-regularised loss term provides an additional soft constraint on the learned weight structure. As a proof of concept, QINN is instantiated in its masked-architecture configuration for the binary classification of software repositories as AI-enabled or classical. On the AIC-199 proof-of-concept dataset, the proposed QINN attains a cross-validated AUC of 99.47% (±1.18%), recall of 100.00% (±0.00%), and F1-score of 99.02% (±1.34%) under QFD-informed structural masking, outperforming a non-learned QFD scoring baseline by 37.37 percentage points in recall and exceeding a cross-validated Random Forest ensemble on AUC by 2.47 percentage points (W = 0, p < 0.05), while producing explanations at three QFD-grounded levels—feature salience, named Technical-Evidence activations, and per-criterion quality requirement scores—that align directly with the EU AI Act documentation obligations. Validation on larger, independently curated datasets and sensitivity analysis of the QFD elicitation process are identified as priorities for future work. A domain-general seven-phase application protocol is provided. Full article

Resource-constrained edge processors deployed on unmanned aerial vehicles and wearable platforms require compact, drift-robust gas classification models for a range of environmental and security monitoring applications, including CBRN-motivated scenarios. Existing approaches rely on server-grade architectures incompatible with edge-board-scale deployment, or on classifiers that chemically degrade severely under long-term sensor drift. Each UCI gas class was mapped to a CBRN behavioral category based on physicochemical analogy (molecular functional group, vapor pressure, and metal-oxide semiconductor (MOS) cross-sensitivity pattern), following established precedent. Analyzed were Ammonia (NH₃), Acetaldehyde (CH₃CHO), Acetone ((CH₃)₂CO), Ethylene (C₂H₄), Ethanol (C₂H₅OH), Toluene (C₆H₅CH₃). We propose herein an end-to-end pipeline integrating a novel 1-D convolutional neural network with depth-wise separable convolutions (LiteSensor-Net), INT8 post-training quantization, structured magnitude pruning, and a knowledge-distillation domain-adaptation module (KD–DM) for sensor drift compensation. Using the UCI Gas Sensor Array Drift Dataset (13,910 measurements; 16 metal-oxide sensors; six analyte gases; a 36-month work span). LiteSensor-Net achieved accuracy = 92.63 ± 2.02%, macro-F1 = 0.898, model size = 5.99 kB INT8 pruned, inference latency = 6.3 ms, RAM footprint = 31.7 kB, and energy per inference = 0.04 mJ (all metrics on Raspberry Pi 4B, ARM Cortex-A72). Under chronological forward-chaining evaluation, KD–DM–20 achieved 47.91 ± 18.79% mean accuracy over Batches 2–10, representing a +9.25 pp improvement over uncompensated NC (38.66%). A six-metric benchmark framework—accuracy, macro-F1, model size, inference latency, RAM footprint, and energy per inference—is introduced to standardize edge-AI gas classifier evaluation. The proposed pipeline provides an open-source, deployable foundation for edge-class gas classification systems, with CBRN detection as a motivating application. Full operational validation on certified chemical simulants remains as future work. Full article

In the context of Kolmogorov complexity, the complexity of an object can be characterized by the length of the shortest algorithm required to describe or compute it. In condensed matter systems under equilibrium and nonequilibrium conditions, macroscopic properties distinct from elementary ones can emerge from large numbers of particles. Such a large system size allows system properties to change as control parameters change, producing phase transitions. Inspired by this analogy, it is natural to consider that an optimized computing system may undergo a transition from an initially random organization to a functionally organized state. In this paper, by adopting a specific multi-task setting based on a typical dynamical system, we propose an evolutionary echo state network that realizes the functional differentiation of a random neural network into two subnetworks—one specialized for memory and the other for program execution. The model suggests a minimal neural mechanism for dynamic processes that extract rules embedded in input sequences and store information over short or long time scales. Because the proposed evolutionary model is driven by constraints that jointly reduce task errors and structural redundancy, the resulting network can be regarded as an optimized descriptor of memory and program functions. To clarify the relationship between the proposed model and algorithmic complexity, we introduce Lempel–Ziv-based complexity indices for both network structure and node-wise reservoir dynamics. Although neither density, effective spectral radius, nor the Lempel–Ziv-based complexity indices were prescribed or optimized in the eESN, the evolved network exhibited structural complexity comparable to conventional ESNs in the corresponding region and slightly higher dynamical complexity. These results suggest that the advantage of eESN is not attributable to a direct maximization of complexity itself, but rather to the evolutionary organization of complexity into a functionally differentiated reservoir that supports both memory-like retention and program-like rule extraction. Full article

This paper presents a behavioral-level, pre-silicon analytical co-design framework for memristor read–write interfaces, intended to establish closed-form design rules that subsequently guide SPICE-level and silicon-level realizations. Memristor-based neural hardware requires interfaces that can program resistance states efficiently while suppressing read disturbance, yet existing designs typically rely on empirical tuning without closed-form analytical rules. We close this gap by deriving a single closed-form operating-window inequality $(v_{\mathrm{on}}<V_{\mathrm{rd}}<v_{\mathrm{off}},V_{\mathrm{wr}}\ge V_{\mathrm{wr}}^{min}\left(T_{\mathrm{wr}}\right))$ from the VTEAM state equation, embedding it in an Energy–Delay–Accuracy (EDA) cost function, and validating the resulting parameter set hierarchically up to MNIST-scale inference. The main finding is that this analytically derived parameter set simultaneously achieves a 96.08% set-cycle energy saving and 90.6% MNIST top-1 accuracy (1.2% below software baseline) under realistic D2D/C2C variability, with every measured number agreeing with its analytical prediction within 2%. The framework is instantiated with a two-phase over-threshold-write and sub-threshold-read timing strategy together with a mutually exclusive PMOS-NMOS path-isolation topology, evaluated through behavioral-level MATLAB simulation under linear-drift and VTEAM models. Behavioral simulation confirms each analytical bound within 2%: a 13.78× resistance window with $\le 0.008\%$ cycle-to-cycle drift, 5.01% read-current CV, and 30.94%/96.08% Reset/Set energy savings versus a no-separation baseline. Transistor-level non-idealities (slew rate, charge injection, RTN, retention aging, peripheral overhead) are bounded analytically; full SPICE/silicon validation is identified as immediate follow-up work. These results establish a reusable, analytically grounded reference design that bridges memristive device modeling, circuit-level interface implementation, and neural network-level usability. Full article

The standard first-level trigger in the Pierre Auger Observatory surface detectors (data analysis in FPGAs immediately after digitization in ADCs) was developed when FPGAs were relatively simple and expensive. Thus, the algorithms developed in the 1990s are relatively simple. Substantial progress in electronics now allows the implementation of very sophisticated mathematical algorithms in very efficient systems and relatively inexpensive FPGAs. A neural network was recently developed as an alternative trigger for recognizing neutrino-induced showers, providing relatively high efficiency and allowing signal profiles from Auger photomultiplier tubes of water-Cherenkov detectors originating from atmospheric showers induced by high background neutrinos to be distinguished from other showers. The chemical composition of ultra-high-energy cosmic rays (UHECR) is complex and still not fully known. Additional tools for online, real-time analysis of potential chemical composition could help address this problem. We simulated a large dataset using the CORSIKA package (for simulating the development of extensive air showers in the atmosphere) and OffLine (for generating Cherenkov radiation in surface detectors and digitizing photomultiplier signals in an analog-to-digital converter). These data served as input to a neural network (using MATLAB tools) that attempted to identify the type of initiating particle. Ultimately, the neural network was implemented on an Arria 10 FPGA to generate real-time neural network triggers directly on the pampas in the surface detector. Both simulations and measurements on the Arria 10 development kit confirmed a high degree of reliability. Full article