FORMULA3 NSF–RCF VINES

Research

Research

FR3 is a strong candidate band for 6G, but indoor channels change constantly as people move, rooms fill, and furniture shifts. FORMULA3 develops the measurements, models, and AI needed to keep service consistent through those changes.

The challenge

Great spectrum, unpredictable rooms

The upper mid-band (FR3, 7–24 GHz) offers hundreds of MHz of spectrum and penetrates walls better than mmWave. That makes it attractive for indoor and outdoor-to-indoor (O2I) 6G services such as immersive AR/VR, tele-surgery, and holographic communication. FR3 links can still be blocked by people, furniture, and changes to a room's layout.

Programmable wireless environments (PWEs) address this by placing controllable surfaces on walls to steer signals around obstacles. But most current control is reactive. It responds after a link fails, which causes latency spikes and throughput drops that emerging applications cannot tolerate.

What dynamics do to channels

  • General channel models break down. In our earlier mmWave studies, fitted statistical models were off by 12–18% once users moved and rooms changed. Similar effects are expected in FR3.
  • Channels become sparse. Outages come and go over time, and many surface elements are shadowed at any given moment.
  • Channels are non-stationary. Channel statistics shift with occupancy, with each user's mobility stage (entering, wandering, leaving), and with the room's layout.
  • So AI struggles to generalize. In our tests, conventional models lost more than 50% of their performance when conditions differed from their training data.

The FORMULA3 approach

Proactive, physics-informed, and context-aware

FORMULA3 is a Framework for Optimal Resource Management Using GeneraLizable AI in FR3 dynamic PWEs. It rests on three ideas.

01 · Proactive control

Act before the outage

Predict future channel gains at every surface, drop the shadowed ones from consideration, and reconfigure ahead of time. Fewer interruptions mean lower latency and less signaling.

02 · Physics-informed AI

Learn with the physics

Build propagation knowledge from validated channel models into the training loss, and densify sparse channel data, so predictions stay accurate with only a few channel sensors.

03 · ISAC context

Sense the room, then decide

Infer occupancy, user mobility, and layout changes from the wireless channel itself. Use that context to pick a specialized model, or to trigger continual learning when the room changes significantly.

Hybrid surfaces in FR3

FORMULA3 combines three kinds of programmable surfaces whose strengths complement each other, balancing passive and active control and adapting to line-of-sight (LoS) and non-line-of-sight (NLoS) transitions.

RISReconfigurable intelligent surfaces: nearly passive beam steering with low power use.
SIMStacked intelligent metasurfaces: multi-layer, active wavefront shaping and beamforming.
MRAMetallic (Fresnel) reflector arrays: low cost and versatile across frequencies.

How the work fits together

Research thrusts

Channel data and models from T1 feed the AI of T2 and T3. T4 validates everything against measurements.

T1

Channel characterization and dataset generation

No public datasets or validated channel models yet exist for FR3 hybrid PWEs that capture near-field effects, hardware constraints, and real-world dynamics. T1 builds both.

T1.1

FR3 hybrid PWE channel characterization

A measurement campaign on fabricated FR3 RIS prototypes and in-house metallic reflector arrays, using CST simulations, anechoic-chamber measurements, and real indoor and O2I deployments. An automated, closed-loop system with motorized positioners sweeps angles while cycling surface configurations. It captures reflection, beam patterns, polarization, inter-element coupling, nonlinear hardware behavior, and near-field effects.

Lead: Ekin (Texas A&M), Rappaport (NYU)

T1.2

Simulation platform for dynamic data

A simulation suite connects CST Studio and NVIDIA Sionna through a shared 3D environment model and custom Python middleware. It produces large datasets with multi-user mobility, changing occupancy, and layout changes. Every sample carries an environment label that T3 uses for context awareness.

Lead: Ekin (Texas A&M)

T2

AI-native proactive resource management and hybrid surface control

Physics-informed AI that builds FR3 channel knowledge into training, so the network can predict, prune, and configure surfaces before problems occur.

T2.1

Surface channel prediction

A physics-informed graph neural network predicts channel gains across RIS tiles and reflector plates from only a few channel sensors. A learned sensor placement balances accuracy against energy use, and a densification step handles temporal sparsity.

Lead: Ismail (Tennessee Tech)

T2.2

Exclusion of shadowed surfaces

Shadowed elements add overhead without helping coverage. An exclusion score accounts for both blockage and LoS/NLoS interference, and an adaptive reinforcement-learning policy decides how often to update it.

Lead: Ismail (Tennessee Tech)

T2.3

User assignment and surface configuration

LoS and reflected signals can interfere destructively. An actor-critic policy assigns users to surfaces and configures SIM, RIS, and MRA together to minimize interference and outage while limiting energy use.

Lead: Basar (Tampere)

T3

Enhancing AI generalization via ISAC-enabled context awareness

A model trained for one occupancy level can lose much of its accuracy at another. T3 aims to keep degradation under unseen dynamics below 2%.

T3.1

Context-aware ISAC framework

A three-step pipeline of sensing, context inference, and model selection. It identifies occupancy level and mobility stage from the channel and invokes a model trained for those conditions. Sensing and resource management are optimized end to end, with environment-aware configuration codebooks.

Lead: Wichman (Aalto)

T3.2

Continual learning under concept drift

When furniture moves or a room is rearranged, the system detects the drift, trains a replacement model in shadow mode with transfer learning, and swaps it in once it outperforms the old one, without interrupting service.

Lead: Ismail (Tennessee Tech)

T4

Validation and evaluation

A unified simulation and measurement platform benchmarks every model across realistic dynamic environments, including mobility, occupancy, and layout changes.

T4.1

Channel models and data

FR3 testbeds at about 7 GHz (Texas A&M) and 15 GHz (Tampere) provide ground truth for the channel models and for statistical benchmarking of the synthetic data.

Lead: Ekin (Texas A&M), Rappaport (NYU)

T4.2

Proactive resource management

Prediction error, delay, complexity, savings in overhead and power, outage, QoS, load balance, and convergence.

Lead: Basar (Tampere)

T4.3

AI adaptation strategies

Context-classification accuracy, inference latency, recovery time after drift, and robustness of static versus adaptive models.

Lead: Wichman (Aalto)

T4.4

End-to-end FORMULA3

Does the full system deliver multi-Gbps throughput and sub-millisecond latency consistently across dynamic scenarios? QoS satisfaction, stability, and adaptation efficiency.

Lead: Ismail (Tennessee Tech)

VINES research vectors

Where FORMULA3 fits in VINES

VINES (Verticals-enabling Intelligent Network Systems) is a joint program of NSF and the Research Council of Finland. FORMULA3 addresses three of its research vectors.

A1

Enabling emerging verticals

  • Multi-Gbps throughput
  • Sub-millisecond latency
  • Indoor and O2I connectivity
  • Reliable QoS under dynamics
B1

AI-native design

  • Physics-informed AI
  • Predictive resource management
  • Generalizable AI
  • Continual adaptation
B3

Integrated sensing and communication

  • Occupancy sensing
  • Mobility awareness
  • Layout and context detection
  • Adaptive network control

Milestones

Timeline

The planned schedule in half-year periods, from the October 2026 start to September 2029.

Task Year 1 Year 2 Year 3
H1H2H1H2H1H2
T1.1FR3 hybrid testbeds and channel characterizationEkin (TAMU), Rappaport (NYU) · FR3 hybrid PWE testbeds at Texas A&M and Tampere; channel data and models
active
active
T1.2Dynamic simulation platform and data generationEkin (TAMU) · Dynamic channel-gain data and models with environment labels
active
active
T2.1Energy-efficient surface channel predictionIsmail (Tennessee Tech) · AI model for surface channel prediction
active
active
T2.2Dynamic exclusion of shadowed surface elementsIsmail (Tennessee Tech) · Exclusion strategy balancing outage, signaling, and energy
active
active
T2.3Interference-aware user assignment and surface configurationBasar (Tampere) · Assignment and configuration policies that improve QoS
active
active
T3.1ISAC framework for context-aware AI selectionWichman (Aalto) · ISAC-based framework for adaptive AI-driven resource management
active
active
T3.2Continual learning under concept driftIsmail (Tennessee Tech) · Transfer-learning-based continual adaptation framework
active
active
T4.1Validate channel models and generated dataEkin (TAMU), Rappaport (NYU) · Statistical benchmarking of synthetic vs. measured data
active
active
active
T4.2Evaluate proactive resource managementBasar (Tampere) · MSE, delay, QoS, overhead, and convergence metrics
active
active
active
active
T4.3Evaluate ISAC and continual learningWichman (Aalto) · Accuracy, latency, recovery, and robustness metrics
active
active
active
T4.4End-to-end FORMULA3 evaluationIsmail (Tennessee Tech) · Throughput, latency, and QoS-consistency metrics
active
active

H1 runs October to March and H2 runs April to September. Hover a bar to see its dates.

Infrastructure

Testbeds and tools

Hardware

Hybrid FR3 testbeds

Fabricated FR3 RIS prototypes and in-house metallic reflector arrays (7–15 GHz), deployed in office and lab settings for indoor and O2I measurements.

Measurement

Anechoic chambers and automation

Chamber measurements at Texas A&M (about 7 GHz) and Tampere University (about 15 GHz). An automated pan-tilt and rail system sweeps angles while cycling surface configurations. Channel sounding uses vector network analyzers and software-defined radios.

Simulation

CST–Sionna suite

Full-wave CST Studio models of surface elements feed NVIDIA Sionna ray tracing through Python middleware, and a REST API serves the data to AI training and testing.

Glossary

Terms used on this site

FR3Frequency Range 3
The 7–24 GHz upper mid-band, between today's sub-7 GHz and mmWave bands, and a candidate band for 6G.
PWEProgrammable wireless environment
A space whose walls carry controllable surfaces, so the network can shape how radio waves travel.
RISReconfigurable intelligent surface
A panel of tiny, tunable elements that redirects incoming signals with very little power.
SIMStacked intelligent metasurface
Several metasurface layers stacked to shape wavefronts and form beams directly in the wave domain.
MRAMetallic reflector array
An array of metal (Fresnel) reflectors. Low cost, and it works across a wide range of frequencies.
ISACIntegrated sensing and communication
Using the communication signal itself to sense the environment, such as how many people are present and how they move.
O2IOutdoor-to-indoor
Links from an outdoor transmitter to users inside a building.
LoS / NLoS(Non-)line-of-sight
Whether there is a direct path between transmitter and receiver, or only reflected paths.
QoSQuality of service
The throughput, latency, and reliability an application actually experiences.
Concept drift
When the data an AI model sees in operation changes from the data it was trained on, for example after a room is rearranged.