The 2026 ISEF global finals have officially concluded. This year, more than 1,700 participants from 67 countries and regions competed at the international final round.

Because of length limits, the first part of our 2026 ISEF winning-project analysis was published earlier.

Now, Part II continues with a closer look at more first-prize projects and the research trends they reveal.

Translational Medical Science (TMED)

This year, the Translational Medical Science category included 94 projects, making it one of the largest categories at ISEF. It showed a strong dual-engine trend centered on AI-assisted diagnosis and new therapeutic strategies. AI and machine learning diagnostic and prediction models accounted for about 35 percent, neuroscience and neurodegenerative disease about 20 percent, and cancer treatment and biomarkers about 18 percent. Drug delivery and nanomedicine represented about 12 percent, while infection and immunity accounted for 8 percent, and regenerative medicine and tissue engineering 7 percent. Together, these projects reflected the full translational spectrum from molecular targets to bedside applications.

Among nearly one hundred projects focused on early cancer detection, neurological disease diagnosis, and new drug development, two first-prize winners stood out. One addressed early stroke recognition in emergency settings using AI analysis of 911 voice and text data. The other focused on retinal disease by identifying a new biomarker, ZNF385D, and proposing the candidate therapeutic AKB-9778.

TMED065 used ordinary 911 call records to enable rapid triage and warning for time-sensitive stroke, directly targeting the pre-hospital emergency time window. TMED062 traced retinal electrophysiological phenomena, especially fixation-related eye movements, back to a specific gene regulatory network and proposed a new small-molecule intervention strategy.

Together, these two first-prize projects captured the core mission of translational medicine: accelerating the conversion of scientific discovery into better solutions for human health.

Overall, TMED065 stood out for its creative use of real-world data and its exceptionally strong feasibility for deployment, directly addressing the clinical principle that time is brain in stroke care without requiring any additional hardware. TMED062 won through a deep progression from phenotype to molecular mechanism and a complete therapeutic proof of concept, representing the classic mechanism-driven path in translational medicine. Together, they illustrate two powerful paradigms in the field: data-driven clinical decision support and mechanism-driven targeted therapy.

Computational Biology and Bioinformatics (CBIO)

The 2026 Computational Biology and Bioinformatics category included 93 projects and showed a clear structure dominated by AI-driven drug development, protein design, multi-omics analysis, and neurological disease modeling.

Research Direction	Approximate Share	Description
Computational Pharmacology	26.9%	Molecular docking, drug screening, antibody design, peptide inhibitors, molecular glue degraders, and personalized treatment strategies
Computational Biomodeling	23.7%	Protein function prediction, cell-state modeling, medical image analysis, disease mechanism simulation, and biophysical process modeling
Genomics	22.6%	Multi-omics integration, spatial transcriptomics, gene expression, CRISPR off-target prediction, and gene optimization
Computational Neuroscience	17.2%	Autism, Alzheimer’s disease, Parkinson’s disease, epilepsy, brain aging, and brain-computer interface modeling
Computational Evolutionary Biology	7.5%	Drug-resistance evolution, pathogen adaptation, ecological dynamics, and directed evolution
Computational Epidemiology	2.2%	Public-health prediction such as epidemic warning and sepsis monitoring

Overall, the dominant trend in this category has expanded beyond traditional sequence analysis toward AI-based drug design, multimodal medical modeling, spatial omics, and disease mechanism inference, with a strong precision-medicine and generative-biology orientation.

One first-prize project, CBIO071, used a Markov chain to predict the failure time of tomato ETI resistance, directly connecting plant immunity, pathogen evolution, and pesticide reduction. The model was clearly structured, used specific parameters, and was validated through temperature conditions, CFU growth ranges, and Roq1 multi-season durability, showing both mechanistic interpretability and practical potential.

Another first-prize project, CBIO018, built an end-to-end diagnostic pipeline combining detection, blood-flow inference, and rupture prediction. It integrated PointNet++, physics-informed neural networks, and Navier–Stokes constraints. With a detection AUC of 0.95 and very low physics loss, the project showed strong innovation and credibility at the intersection of medical imaging, fluid mechanics, and AI.

Restored Poster Information for CBIO018

Project Information	Content
Title	A Multi-Modal Deep Learning Pipeline Integrating Physics-Informed Neural Networks and 3D Image Processing for Accurate Rupture Prediction and Data-Driven Surgical Planning of Cerebral Aneurysms: Year 3
Project Number	CBIO013
Author	Eshan Vipuil
School	West Shore Junior/Senior High School
Location	Melbourne, Florida, United States

Problem

• Invasive endovascular surgery
• Need for selective treatment
  • Treat high-risk aneurysms
  • Avoid unnecessary high-risk procedures

Data Analysis

• Superior detection
• Accurate flow simulation
• Improved rupture prediction

Methodology

1. Geometry-based aneurysm detection trained on the Vascular Model Repository
2. Individualized Navier-Stokes physics-informed neural network
3. Multimodal rupture prediction using geometry, flow, and biological factors

Conclusions

• The detection model automates the clinician’s role in aneurysm identification
• PINN correction accelerates CFD simulation for deriving blood flow
• The prediction model extends beyond clinician capability in surgical planning

Embedded Systems (EBED)

In 2026, the Electronics and Embedded Systems category included 49 projects and showed a technology structure centered on intelligent sensing, edge AI, and low-cost wearable devices.

Technology Direction	Approximate Share	Description
Sensors	49.0%	LiDAR, inertial and pressure sensing, acoustic detection, air-quality monitoring, plant electrophysiology, radiation detection, and assistive sensing devices
Microcontrollers	16.3%	Raspberry Pi, ESP32, FPGA, embedded vision, and automated control systems
Signal Processing	12.2%	Audio, spectrum analysis, radar, medical imaging, and signal recognition
Networking and Data Communications	8.2%	Wireless communication, LoRa, Wi-Fi, and interference-resistant communication
Circuits	8.2%	Analog and digital circuit design
Internet of Things	4.1%	Mainly for environmental monitoring and smart warning systems
Optics	2.0%	Imaging and visual-scanning potential

The first-prize project developed a low-cost stereo-vision 3D scanner for microscopic samples. It embedded binocular cameras into a microscopic optical effector and used multi-angle acquisition, semi-global matching, disparity mapping, and triangulation to generate point clouds for nondestructive 3D reconstruction. The system integrated a Jetson Nano, Teensy control, ROS, and RViz visualization, and could be used for particle structure analysis, biological samples, and material defect inspection. Its advantages included low cost, modularity, and compatibility with robotic arms.

Restored Poster Information for EBED006

Project Information	Content
Title	A Novel Stereo Vision 3D Scanning System for Microscopic Samples
Author	Filip Lajciak
School	Secondary Technical School
Location	Dubnica nad Vahom, Slovakia
Project Number	EBED006

Introduction

• Problem: Current methods rely on expensive laser scanning microscopes, often costing hundreds of thousands of dollars. These devices scan only from a top-down perspective and miss defects or structures visible from other angles.
• Solution: Rotate the entire stereo imaging module around the sample to achieve full angular coverage. Integrate stereo cameras into microscope optics to capture depth maps from a microscopic viewpoint.

Software

• Built on ROS, synchronizing the camera, motor, and processing pipeline into one system
• Disparity is converted into depth via triangulation to generate a point cloud of the scanned surface

Hardware

• Version 1.0: Conventional stereo microscope with integrated stereo camera; image streams controlled by an NVIDIA Jetson Nano
• Version 2.0: A motorized effector rotates around the sample with one optical microscope compared with Version 1.0

Results

• Best depth resolution of 15 micrometers achieved using the motorized optical effector at 200x magnification
• Average deviation from the ideal sphere surface was 0.2 mm
• Use cases included outer defectoscopy, implant metrology, and welding-line inspection

Energy: Sustainable Materials and Design (EGSD)

The 2026 Energy category included 49 projects and showed a balanced structure across bioenergy, solar materials, and thermal energy conversion.

Research Direction	Approximate Share	Description
Biological Process and Design	24.5%	Algal fuels, bioethanol, anaerobic digestion, microbial fuel cells, and biofilm materials
Solar Process, Materials, and Design	18.4%	Perovskites, photovoltaic interfaces, photothermal evaporation, and solar-cleaning strategies
Thermal Generation and Design	16.3%	Thermoelectrics, heat pipes, temperature-difference power generation, and high-temperature conversion
Hydrogen Generation and Storage	14.3%	Green hydrogen catalysis and electrolysis
Wind and Water Movement Power Generation	14.3%	Wind and wave power optimization
Triboelectricity and Electrolysis	8.2%	—
Energy Storage	4.1%	Smaller in number but with potential for materials innovation

The first-prize project addressed real-world fluctuations in wind speed and direction by optimizing 50 airfoil sections of the IEA 15 MW offshore turbine using CST parameterization, a surrogate-assisted genetic algorithm, and an XGBoost fitness function, then validating the design through aero-servo-hydro-elastic simulation. Its strength lay in expanding from single-airfoil optimization to full-turbine blade reconstruction while also considering manufacturability and real offshore conditions. The projected gain was 2,236 MWh per turbine per year, giving it strong application value.

Restored Poster Information for EGSD011

Project Information	Content
Title	Optimization of Wind Turbine Performance: Integrating Evolutionary Based AI-Assisted Genetic Algorithms and XGBoost with Blade Element Momentum and Aero-Servo-Hydro-Elastic Simulations
Author	Janak Vasisht
School	H-B Woodlawn
Location	Arlington, Virginia, United States
Project Number	EGSD011

Research Problem and Objective

• Problem: Wind turbine blade design research often optimizes for narrow wind speeds and angles, producing turbines that are not robust under real conditions.
• Objective: Build a framework that optimizes power output through blade-shape optimization across all environmental conditions.

Design Criteria

• Maximize power performance under realistic conditions
• Maintain manufacturability
• Respect computational-cost limits
• Limit deviations from the original blade shape

Engineering Methodology

• Airfoil Parameterization: Class Shape Transformations to define airfoil geometry
• Surrogate Model Function: Predict airfoil performance across a range of wind angles
• Surrogate-Assisted Genetic Algorithm: Optimize each airfoil across realistic environmental conditions by maximizing the objective function

Results and Performance Analysis

• All 50 optimized airfoils were assembled into a full optimized blade, and all three blades formed the full turbine
• Offshore aero-servo-hydro-elastic simulations evaluated turbine performance
• A regression function was fitted to wind speed and reward-difference data
• Annual revenue and power difference were calculated using hourly wind-speed data

Performance Metrics

• +2,236 MWh per turbine
• +$536,426 per turbine
• +$68.8 million nationally

Interpretations and Conclusions

• The surrogate-assisted genetic algorithm optimized 50 airfoils of the IEA 15 MW Offshore Reference Turbine
• The architecture improved the lift-to-drag ratio of one airfoil by as much as three times
• The optimized turbine produced an additional 2,236 MWh annually under 2025 wind conditions
• Net revenue difference between turbines was $536,426.54
• Structural feasibility was evaluated by simulating load and stress

Engineering Technology: Statics and Dynamics (ETSD)

The 2026 Engineering Mechanics and Mechanical Engineering category included 55 projects and was dominated by aerospace and mechanical system design.

Research Direction	Approximate Share	Description
Aerospace and Aeronautical Engineering	38.2%	UAVs, rockets, airfoil optimization, propulsion systems, morphing wings, and landing gear
Mechanical Engineering	27.3%	Transmission systems, bearings, heat exchange, rehabilitation devices, robotics, and energy-conversion structures
Ground Vehicle Systems	9.1%	Aerodynamics, crash safety, and autonomous driving systems
Industrial Engineering – Processing	9.1%	Manufacturing, refining, drilling, and process optimization
Civil Engineering	7.3%	—
Naval Systems	5.5%	—
Control Theory	Small share	Control modeling
Computational Mechanics	Small share	High-precision simulation in engineering design

HandTalk is a low-cost ASL translation and reproduction platform that combines MediaPipe pose recognition, machine-learning classification, and a 6-DOF robotic arm with a 10-DOF robotic hand to achieve real-time sign-language recognition, imitation, and playback. A Bi-LSTM model achieved about 92 percent accuracy on 27 vocabulary items, showing value for robot-assisted communication, accessible education, and low-resource communities.

Restored Poster Information for ETSD048

Project Information	Content
Title	HandTalk: A Translation System for American Sign Language
Author	Ana Spride
School	Plano East Senior High School
Location	Plano, Texas, USA
Project Number	ETSD048

Engineering Problem and Objectives

• Defined Need: 430 million deaf and hard-of-hearing individuals worldwide and 600,000 in the United States face communication challenges in everyday life
• Engineering Goal: Recognize ASL words in real time, build a 16-DOF robotic arm to reproduce ASL signs, and prioritize ASL performance and cost-effectiveness
• Design Constraints: Easy to use and low-cost components

Project Design

• ASL Recognition: Video frame feature extraction, sign detection, and gesture classification
• Robotic Arm: Construction of a 16-DOF arm

Results

• Neural networks significantly outperformed static models
• Temporal models such as BiLSTM performed best
• Distance-based methods were limited
• Robotic sign reproduction was feasible
• Shoulder-arm drift resolution was still under investigation

Conclusions

• ASL translation worked
• 27 ASL words were trained on 800 videos
• Recognition plus robotic reproduction was achieved
• BiLSTM was the best-performing model
• Hand repeatability reached 96 percent

Applications

• ASL learning feedback tools
• Outreach in schools and public spaces
• Rehabilitation support in training and use modes

Limitations

• Gaps in signer coverage
• Shoulder-arm drift

Future Work

• Stabilization and inverse kinematics
• Expanded vocabulary and signer set
• Hybrid bio-silicon investigation

Environmental Engineering (ENEV)

The 2026 Environmental Engineering category included 88 projects and centered on pollution control, water-resource management, and waste recycling.

Research Direction	Approximate Share	Description
Pollution Control	40.9%	PFAS, microplastics, heavy metals, oil pollution, air pollution, and wastewater contaminants
Water Resources Management	23.9%	Water-quality monitoring, flood management, water conservation, and marine and freshwater ecosystem governance
Recycling and Waste Management	20.5%	Plastics, textiles, agricultural waste, and bio-based recycling
Land Reclamation	8.0%	Soil improvement, vegetation restoration, wildfire recovery, and coastal ecology
Bioremediation	6.8%	Fungal, bacterial, enzymatic, and plant-mediated pollutant degradation

ENEV043 developed an in situ microplastic detection system mounted on a biomimetic sea-turtle robot. By combining ultra-compact digital holography, image-reconstruction algorithms, and AI classification, the system enabled real-time underwater microplastic detection. It achieved sub-10-micrometer resolution and 94 percent accuracy for PS, PE, and PET detection, and was validated through CFD, pool testing, and 15 field tests across 10 bodies of water. Its value lies in low-cost, broad-scale, and environmentally friendly ocean-pollution monitoring.

ENEV044T proposed a crop-protection system combining biopesticides with robotic precision spraying. It first activated plant anti-insect responses using MeJA, then used dsRNA to silence growth-inhibition enzymes, reducing the growth penalty caused by high-dose MeJA, with carbon-dot nanocarriers protecting the RNA. The system also developed a mecanum-wheel autonomous vehicle using SIFT and YOLOv5 to identify pests in real time and spray precisely based on pest quantity, reducing yield loss by about 40 percent compared with manual spraying.

Mathematics (MATH)

The 2026 Mathematics category included 41 projects and showed a structure dominated by probability and statistics, graph theory and game theory, and analysis.

Research Direction	Approximate Share	Description
Probability and Statistics	29.3%	Epidemic prediction, financial networks, time series, neuroscience, and social-data modeling
Combinatorics, Graph Theory, and Game Theory	22.0%	Steiner trees, gerrymandering, routing games, hypergraphs, and combinatorial systems
Analysis	17.1%	PDEs, numerical methods, operator theory, and dynamical systems
Geometry and Topology	14.6%	Persistent homology, virtual knots, projective geometry, and fractal dimension
Number Theory	Small share	Primes, finite fields, group structures, and cryptography
Algebra	Small share	Group structures and cryptographic theory

The first-prize project studied whether complex meromorphic equations of the form f(x) = a can be solved using elementary functions. Its innovation lay in using topological Galois theory to generalize previous unsolvability results for special equations such as tan x − x = a and x^x = a to a much broader class of functions. The result showed that if f′ has infinitely many roots and the corresponding critical values are infinite, then the equation cannot be solved in elementary functions. The project displayed remarkable theoretical depth through the fusion of advanced analysis, group theory, and Galois theory.

Restored Poster Information for MATH006

Project Information	Content
Title	Solvability of Meromorphic Equations in Elementary Functions
Author	Nikola Veselinov
School	Sofia High School of Mathematics
Location	Sofia, Bulgaria
Project Number	MATH006

Research Question

• Elementary functions are finite compositions of addition, subtraction, multiplication, division, exponentials, logarithms, and constants
• For equations of the form f(x) = a, when is x = f⁻¹(a) elementary?

Main Result

• Let f be a meromorphic elementary complex function. If the set of critical values {f(x₀): f′(x₀) = 0} is infinite, then f(x) = a is unsolvable in elementary functions.

Framework

• Closed curves of a around critical values induce permutations of roots that generate the monodromy group
• f(x) = a is solvable whenever the monodromy group is a solvable group

Main Proof

• Relate primitivity of the monodromy group to decomposition f = g₁ ○ g₂ ○ ... ○ gᵣ
• Apply Wielandt’s theorem for primitive infinite permutation groups to prove that at least one quotient gᵢ has an unsolvable monodromy group
• Conclude unsolvability through the infinitude of the set of critical values

Conclusion

• The result generalizes previous unsolvability results to all equations involving meromorphic functions with infinitely many critical values
• The project also proposed a conjecture that if f(x) = a is solvable, then f can be decomposed into depth-1 compositions without nested exponentials or logarithms and with primitive monodromy groups

Physics and Astronomy (PHYS)

The 2026 Physics category included 82 projects and was dominated by astronomy, cosmology, and quantum or computational physics.

Research Direction	Approximate Share	Description
Astronomy and Cosmology	37.8%	Exoplanets, dark matter, black holes, galaxy evolution, gravitational waves, and solar activity
Theoretical, Computational, and Quantum Physics	18.3%	Quantum algorithms, decoherence, general relativity, and physics-informed modeling
Mechanics	11.0%	Fluids, vibration, drag, and rotational dynamics
Condensed Matter and Materials	9.8%	Superconductivity, phase transitions, nanomaterials, and hydrogen-storage materials
AMO	Small share	Optical imaging
Biological Physics	Small share	Biophysical modeling
Nuclear and Particle Physics	Small share	Theoretical depth in particle and nuclear topics

PHYS021 proposed using MCMC to sample the complete configuration space of origami and linkage mechanisms. Its innovation lay in representing mechanisms as graph structures and borrowing ideas from statistical mechanics to build an artificial energy landscape in which rigid configurations correspond to low-energy states. It then combined Metropolis methods, Hamiltonian Monte Carlo, and parallel tempering to cross different motion modes. The framework could verify known analytical solutions, analyze bistability in ladybug wing folding, and extend to inverse design of mechanisms.

PHYS030T transferred Parker Solar Probe near-sun plasma observations into tokamak ICRF heating optimization by building a workflow that included PSP diagnostics, dimensionless mapping, AORSA full-wave calculations, and Gkeyll/BOUT++ analysis. Its key innovation was using natural plasma observations to guide fusion-heating design, increasing optimal ion absorption power by 2.36 times, increasing ion-heating proportion by 2.46 times, and accelerating the process by a factor of 84.

Robotics and Intelligent Machines (ROBO)

The 2026 Robotics and Intelligent Machines category included 66 projects and was dominated by machine-learning-driven perception, decision-making, and autonomous systems.

Research Direction	Approximate Share	Description
Machine Learning	45.5%	Visual recognition, deep learning, reinforcement learning, multimodal models, disaster detection, and medical imaging
Robot Kinematics	28.8%	Drones, mobile robots, biomimetic robots, robotic arms, soft robotics, and navigation systems
Cognitive Systems	13.6%	LLM reasoning, cognitive modeling, AI companions, and explainable intelligence
Control Theory	7.6%	Flight control, path planning, SLAM, and motion control
Biomechanics	4.5%	EMG, brain-computer interfaces, gait, and exoskeletons

ROBO021 developed a low-cost ellipsoidal amphibious wetland-monitoring robot for in situ detection of water-quality indicators such as pH and turbidity. The system integrated a sealed shell, a three-stage gear drive, a sliding balance mechanism, PID control, Kalman filtering, LoRa communication, and K210 edge-vision recognition, allowing stable operation in shallow water, mud, and dense vegetation. Experiments showed trajectory deviation of no more than 2.6 percent, latency under 92 milliseconds, and a cost of about $380, making it suitable for large-scale wetland ecological monitoring.

ROBO035 developed a low-cost CTIS hyperspectral imaging system and a physics-aware self-supervised AI reconstruction framework. Its core innovation, the PASS Transformer, embedded imaging physics such as diffraction and point-spread functions into attention mechanisms, enabling restoration of high-dimensional spectral information without requiring full real hyperspectral ground truth and reducing hallucination and reconstruction ambiguity. The system was validated in food-allergy detection, plant-stress monitoring, and environmental pollution analysis, combining low hardware cost with high algorithmic accuracy.

Restored Poster Information for ROBO035

Project Information	Content
Title	Decoding Light: Physics-Aware Self-Supervised AI for Low-Cost High-Fidelity Hyperspectral Imaging
Author	Michael Hua
School	Cranbrook Kingswood School
Location	Bloomfield Hills, Michigan, USA
Project Number	ROBO035

Research Problem

• Hyperspectral imaging captures both spatial and spectral information and reveals material properties invisible in RGB
• Real-world HSI systems are expensive and require extensive calibration

Research Goal

• Develop a low-cost CTIS system with high-fidelity reconstruction for real-world deployment

Hardware Design

• Hardware cost under $500
• Portable, 3D-printable, and open-source

Software

• PASS Transformer, a physics-aware self-supervised transformer network
• Integrates diffraction, point-spread function, and other imaging physics into the attention mechanism

Findings and Results

• Real CTIS measurements under incoherent light were captured in the prototype for training and testing
• PASS Transformer outperformed existing methods in reconstruction quality

Applications and Conclusions

• Precision agriculture
• Food allergy detection
• Environmental monitoring
• The PASS Transformer enabled fast, high-fidelity, nondestructive hyperspectral analysis with strong real-world potential

Software Systems (SFTD)

The 2026 Software Systems category included 66 projects and was dominated by algorithmic innovation, human-machine interaction, and cybersecurity.

Research Direction	Approximate Share	Description
Algorithms	37.9%	Deep learning, reinforcement learning, optimization, generative models, clustering, physics-informed modeling, and prediction systems
Human/Machine Interface	24.2%	Gestures, eye tracking, voice, AR, AAC communication tools, and accessible interaction
Cybersecurity	15.2%	Vulnerability detection, deepfake recognition, watermarking, anti-fraud, post-quantum cryptography, and random-number security
Mobile Apps	9.1%	Application platforms
Online Learning	6.1%	Educational support tools
Databases	Small share	Data management
Languages and Operating Systems	Small share	Programming environments and system-level security

SFTD001, called M.A.N.T.I.S, is a fully computational underground-imaging system using natural atmospheric muons, Geant4 simulation, machine learning, and multi-source remote-sensing data to achieve remote subsurface structure recognition without excavation or on-site detectors. The project integrated 32 data sources including InSAR, GRACE, and Sentinel-2, and trained on more than 8 million simulated datasets. It has potential applications in disaster response, infrastructure planning, mining, and national security, with advantages of low cost, noninvasiveness, and minimal environmental disturbance.

Restored Poster Information for SFTD001

Project Information	Content
Title	M.A.N.T.I.S: Muon Analysis for Non-Invasive Tomography Image Simulation
Author	Zack O’Leary
School	Clongowes Wood College
Location	Kildare, Ireland
Project Number	SFTD001

Engineering Problem and Objectives

• Problem: Current subsurface imaging techniques are costly, invasive, inaccurate, and often require on-site equipment
• Solution: M.A.N.T.I.S is a full-stack machine-learning subsurface imaging system that simulates muon interactions and identifies structures with no user input beyond location coordinates

Procedures

1. Data acquisition and ingestion
2. Physics-based forward modeling
3. Inversion and knowledge extraction
4. Anomaly detection
5. Validation and data synthesis

Data Analysis and Results

• 15 custom Python scripts using NumPy, Matplotlib, and Open3D
• Parsing of millions of GEANT4 muon trajectories
• Topological clustering to isolate meaningful signal patterns
• Geometric reconstructions for the SURF mine
• System output evolved from low-definition to high-definition imaging at about 5 to 6 meters

Validation Pipeline

• Coordinates plus public data to prior generation to muon simulation in GEANT4 to inversion and uncertainty to survey output

Conclusions and Implications

• Lower risk and cost, with no environmental disturbance
• Enables remote imaging
• Validation at the SURF mine and other sites demonstrated accuracy even for highly complex underground geometry
• Technology stack included Python, C++, GEANT4, React, Docker, Google Earth Engine, Neo4j, NumPy, and Open3D

SFTD022T, called ExpressBuddy, is an AI practice companion for children with communication disorders. Through a role-reversal mechanism in which children teach Pico how to respond to bullying, exclusion, and other social situations, it trains pronunciation, turn-taking, and social reasoning. The system supports low-compute Chromebooks and integrates local speech denoising, VAD segmentation, facial-expression recognition, and imitation feedback. Pilot results showed a 56.8 percent increase in daily verbal output and a 41.2 percent reduction in AI prompting, suggesting real value as a speech-therapy support tool.

Restored Poster Information for SFTD022T

Project Information	Content
Title	ExpressBuddy: An AI-Powered Companion for Children with Communication Challenges
Authors	Sanjay Javangula, Soham Shekhar
School	Bentonville High School
Location	Arkansas, USA
Project Number	SFTD022T

Engineering Problem and Objectives

• 1 in 14 U.S. children has a communication disorder, but 40 percent receive no intervention services
• Children receive only 1 to 3 hours per week of therapy support, leaving more than 165 hours without guidance
• Children with speech disorders often face social and emotional difficulties and are twice as likely to be bullied
• Goal: Empower children with speech disorders to build real-world speech confidence through safe, judgment-free AI conversation practice

Results

• Pilot testing showed technical feasibility and strong real-world impact
• Speech-language pathologist feedback described Pico as flexible for therapy targets
• Average student rating was 4.5 out of 5
• All participants strongly agreed they would recommend ExpressBuddy to a friend

Interpretations and Conclusions

• ExpressBuddy bridges the gap between therapy sessions and reinforces clinical skills effectively
• Its multimodal approach combining facial tracking and voice may engage the brain’s mirror-neuron system and create a safe parasocial bond
• For safety and privacy, raw audio is discarded after processing and only text transcripts are stored
• Future work includes scaling to serve millions of students and expanding social scenarios

Technology Enhances the Arts (TECA)

In 2026, the TECA-related category included 20 projects. Human Information Exchange accounted for about 45 percent and focused on sign-language translation, speech-disorder support, smart glasses, hearing-impaired communication, and cognitive care. Music and Image Manipulation made up about 25 percent, including music practice, humming recognition, poetry rewriting, and music therapy. 3D Modeling and Display Technology each accounted for about 10 percent, reflecting trends in 3D reconstruction, smart glasses, and haptic interaction. Games and Engineering Effects each accounted for about 5 percent.

TECA008, called HARMONI, is an AI-driven music-therapy platform that uses multimodal emotion recognition through facial expression, vocal prosody, physiological signals, and psychological scales. It combines retrieval-augmented personalization, conversational support, image generation, and music generation to provide adaptive intervention. Experimental results showed emotion-recognition accuracy of 88 percent and significant reductions in anxiety, stress, and heart rate, highlighting its value as a low-cost, scalable, multilingual mental-health support tool.