At Engiscent, we develop Physics-Enhanced Neural Networks (PENN) — hybrid AI models that blend neural architectures with the governing laws of physics. This fusion allows us to create AI that is not only powerful, but also interpretable and physically grounded.
Our Key Research Domains
🎧
AI for hyperspectral and
multispectral imaging.
🌈
AI for sound-based diagnostics
and monitoring.
🧠
AI for real-time brain
signal analysis.
📡
AI for integrating sensor
data and predictions.
We apply deep convolutional and Transformer-based architectures enhanced with physical constraints (e.g., radiative transfer models, atmospheric absorption coefficients). PENN aligns these models with Maxwell’s equations and optics principles, allowing precise calibration and interpretation of reflectance data.
Technical Approach
Spectral unmixing with physics-informed CNNs
Separating mixed spectral signatures for accurate material identific
Bidirectional Reflectance Distribution Functions
Modeling light interactions to improve surface property estimation.
Atmospheric correction modules
Eliminating distortions from atmospheric interference in remote sensing data.
Time-series fusion with attention
Enhancing temporal consistency in spectral analysis.
Applications
Vegetation health
diagnostics
Identifying stress factors in crops and forests using hyperspectral imaging.
Soil mineral composition
mapping
Modeling light interactions to improve surface property estimation.
Urban heat island
detection
Monitoring heat absorption and retention patterns in cities.
Underwater bathymetric
mapping
Deriving seabed topography from satellite imagery for marine studies.
We use deep learning architectures like spectrogram-based CNNs and phase-aware neural networks, informed by physical wave propagation models such as Helmholtz and Navier-Stokes. Our models are optimized for real-time, edge-device inference in industrial and agricultural settings.
Technical Approach
Spectrogram and wavelet based CNNs trained with physics priors
Extracting spectral patterns for improved signal analysis.
Sound event localization
using phase-aware neural
networks
Determining sound sources with high spatial accuracy.
Vibration modeling using material-specific resonant frequency parameters
Identifying faults through structural resonance patterns.
Real-time
classification
on edge audio sensors
Enabling low-latency AI-driven sound recognition.
Applications
Fault prediction
in rotating
machinery
Detecting wear and imbalance before failure occurs.
Livestock health monitoring via vocalization patterns
Identifying stress and illness through animal sounds.
Acoustic emission detection in composite materials
Identifying microcracks and structural weaknesses.
Detection of pests or insects through micro-acoustic signatures
Monitoring infestations via subtle sound cues.
We interpret EEG signals using physics-inspired priors, such as neural mass models (e.g., Jansen-Rit, Kuramoto), to enhance accuracy across brainwave bands. AI architectures like temporal CNNs and LSTM hybrids improve signal decoding, even with low SNR and high user variability.
Technical Approach
Temporal Convolutional Networks (TCNs) for real-time brain state tracking
Capturing neural dynamics for cognitive assessment.
Cross-frequency coupling detection with complex-valued neural networks
Identifying interactions between brainwave frequencies.
Phase synchrony analysis with LSTM-physics
hybrids
Tracking neural coherence for cognitive studies.
Domain adaptation for inter-subject variability reduction
Enhancing brain signal models across different users.
Applications
Adaptive therapeutic games for focus & relaxation traning
Personalizing cognitive therapy for users.
Behavioral trait analysis in HR and gaming
contexts
Identifying decision-making patterns in real time.
Brain-computer interface (BCI) for cognitive load estimation
Measuring mental effort through neural signals.
Fatigue detection in drivers and pilots
Preventing accidents through real-time exhaustion monitoring.
We fuse asynchronous sensor data using PENN models informed by physical laws like Fourier’s and Darcy’s. These constraints enhance anomaly detection and forecasting. Our systems incorporate graph neural networks and attention-based architectures for robust time-series modeling.
Technical Approach
Graph Neural Networks (GNNs) for spatial correlation modeling
Separating mixed spectral signatures for accurate material identific
Spatiotemporal signal de-noising using physics-aware autoencoders
Modeling light interactions to improve surface property estimation.
Sensor drift correction using Kalman Filter-enhanced LSTMs
Eliminating distortions from atmospheric interference in remote sensing data.
Temporal Fusion Transformers for sequence prediction
Forecasting trends from time-series sensor data.
Applications
Smart farming via weather, soil, and moisture sensor fusion
Optimizing crop yield through environmental monitoring.
Cyber intrusion detection based on electric and thermal signal anomalies
Identifying hacking attempts via power fluctuations.
Water quality tracking using sensor grids and river flow physics
Monitoring pollution levels for environmental safety.
Remote mining and
geology monitoring
Enhancing resource exploration with AI-driven insights.
Business upon established technology is built by founders who are invited by Engiscent to head spin-offs and become co-founders.
The prospects of AI and IoT technology development effect on the carbon market regulation
Read the paper
Prospects for the use of 5G standards for the protection and monitoring of power plant equipment
Read the paper
An Automated Hardware-Software Module Monitoring Acheta Domesticus Population at Breeding Facilities
Read the paper
Renewable Energy Ventures
Read the paper
Pure Solar Power Initiative
Read the paper
Solar Sky Energy Solutions
Read the paper