Préserver le sens.
A structure, a specimen and a measurement are different scientific objects. Keep conditions, units, method and uncertainty attached to every observation—even when the data are transformed.
CETTE / Intelligence scientifique
Matter.
Reimagined.
De la matière à la découverte.
Meet ok_computer2.0. Our scientific AI architecture connects materials, properties and processes with the evidence needed to decide what comes next.
Brooklyn, New York Une ambition scientifique.
possibilité.
01 / L’architecture
A sharper
question.
ok_computer2.0
What could this material become—and what would we need to know?
ok_computer2.0 is designed to connect proprietary scientific observations with specialised models, physical simulations and experimental learning. The goal extends beyond finding a crystal: discover useful combinations of properties and the processes that make them possible.
Structure. State. Temperature. Processing history. Measurement method. These details determine what a property means. The architecture keeps them attached as a question moves from atoms to microstructure, fields and devices.
Prédire. Vérifier. Apprendre.
Keep predictions, calculations and measurements distinct.
Explorer la matière.
Explore crystal structures, recipes and processing routes against a defined question. Track stability, useful properties and the evidence required to make a candidate worth pursuing.
Relier les échelles.
Connect atomistic models to microstructure, physical fields and devices. Carry uncertainty and model discrepancy through each justified change of scale.
Choisir la suite.
The next useful action may concern a candidate, a competing phase, a control or a repeat measurement. Compare its expected contribution to the decision with its cost and feasibility.
Les réseaux & les méthodes
Specialists. In concert.
Explore the neural networks and scientific methods across nine areas of the engine. Open an area to see each model’s architecture, inputs, outputs and role in the research loop.
Scientific area / 01
Learn the physical structure.
CETTE-ÉQUIAT · ÉQUIAT-LR · ÉQUIAT-MF
Equivariant message-passing models respect the transformation rules of atomistic quantities. Energy-based variants connect predicted energy to forces and stress through derivatives.
Periodic graphs · equivariant representations · energy-based learning
Neural network family
CETTE-ÉQUIAT
Conservative equivariant atomistic learning
ÉQUIAT learns how the local environments of atoms contribute to the energy of a material. Periodic crystal graphs connect species through geometric displacements; radial features and angular representations carry the information needed to preserve physical transformation rules. The scalar energy is differentiated to obtain atomic forces and, with a defined strain convention, stress. This connects atomistic inference to structural relaxation, mechanical response and dynamical calculations within the model’s supported chemistry and conditions.
Equivariant message-passing neural energy model with radial functions, angular representations and a scalar energy readout.
- Inputs
- Atomic species and positions · Lattice and periodic boundary conditions · Charge, spin or method context required by the task
- Outputs
- Potential energy · Forces from energy derivatives · Configurational stress under the selected strain convention
Within the engineProvides the atomistic energy surface used by relaxation, relevant dynamics and derivative-based property workflows; supplies compatible information to crystal generation and scale-transfer modules.
Neural network family
CETTE-ÉQUIAT-LR
Atomistics with explicit long-range interactions
ÉQUIAT-LR extends the atomistic family to problems where interactions beyond a local geometric neighborhood matter. The model combines learned short-range contributions with a declared long-range energy treatment. A charge-equilibration or polarizable route can represent environment-dependent response when the physical formulation and data support it. The full energy, including geometry-dependent long-range terms, determines forces and stress.
ÉQUIAT energy backbone coupled to a defined electrostatic, dispersion or polarizable contribution; charge-response terms enter the total-energy derivatives.
- Inputs
- ÉQUIAT structural inputs · Total charge and applicable electrostatic boundary conditions · Parameters and reference decomposition for the supported long-range interaction
- Outputs
- Total energy with an explicit long-range contribution · Consistent forces and stress · Charge or polarization response for the selected physical formulation
Within the engineExtends the same atomistic interfaces to qualified bulk, slab, charged-system or interface domains, using a separate boundary treatment for each supported setting.
Neural network family
CETTE-ÉQUIAT-MF
Method-conditioned atomistic predictions
ÉQUIAT-MF connects calculations made under different, explicitly identified scientific methods. Its method-conditioned formulation can share structural information while keeping reference energies, observables and correction conventions distinct. Paired calculations on common states support the learned relationship between methods. This lets the engine use complementary computational evidence while retaining the meaning of each prediction.
Shared atomistic representation with method-conditioned heads or a learned discrepancy relationship between compatible reference methods.
- Inputs
- Atomic state and relevant physical conditions · Method identity and reference convention · Compatible labels and linking calculations on shared states
- Outputs
- Method-specific energy or other supported observable · Estimated difference between compatible method outputs · Uncertainty for the conditioned prediction
Within the engineWorks with the method-compatibility service and action selector to choose which calculation can most improve a material decision.
Scientific area / 02
Propose what to investigate.
CETTE-SYMGÈNE
Constrained generative models explore periodic crystal composition and structure. A proposal carries chemistry, cell and task constraints into independent verification.
Conditional generative modelling · symmetry and chemistry constraints
Neural network family
CETTE-SYMGÈNE
Conditional crystal generation
SYMGÈNE explores compositions and periodic structures conditioned on useful properties and permitted chemistry. The generator represents species, lattice and fractional coordinates together, with constraints on cell volume, stoichiometry and optional crystal symmetry. Symmetry-conditioned branches can organize proposals by compatible space groups and Wyckoff multiplicities; other branches allow symmetry breaking and lower-symmetry candidates. Proposed structures then enter independent relaxation and property evaluation.
Periodic composition and structure generation through conditional diffusion or flow-based model routes, with explicit crystal constraints.
- Inputs
- Permitted elements and composition constraints · Target property conditions · Cell-size range and optional symmetry specification
- Outputs
- Proposed compositions, lattices and atomic coordinates · Generator seed, conditioning and proposal lineage
Within the engineConnects atomistic prediction to SYNTHÉVAL process assessment and the engine’s independent relaxation, phase and property evaluation routes.
Scientific area / 03
Connect a material to a route.
CETTE-MIXPROC · CETTE-SYNTHÉVAL
Recipe and process models connect ingredients, sequence and manufacturing conditions to the evidence for making a desired material.
Process-sequence models · constrained proposals · conditional feasibility
Neural network family
CETTE-MIXPROC
Mixture and processing design
MIXPROC searches the joint space of ingredients, proportions and manufacturing history. Its representation separates discrete ingredient choices from continuous fractions and ordered process steps. Temperatures, durations, atmosphere, moisture and equipment limits are represented alongside composition. This allows the engine to explore how composition and preparation combine to influence material performance.
Conditional composition and process-sequence models within a typed ingredient and equipment grammar.
- Inputs
- Ingredient identities and fraction basis · Composition, particle or precursor descriptors · Process steps, temperatures, durations and equipment limits · Target properties and operating constraints
- Outputs
- Mixture proposals · Ordered processing or curing sequences · Candidate preparation windows and associated uncertainty
Within the engineFeeds SYNTHÉVAL route assessment, CINÉTA reaction dynamics and MÉSOLIEN composition-to-microstructure mappings, including the concrete and reactive-material programs.
Neural network family
CETTE-SYNTHÉVAL
Condition-aware synthesis feasibility
SYNTHÉVAL evaluates how a proposed material could be prepared within a real precursor and equipment envelope. It relates candidate identity to prior reactions, precursor characteristics, mixing, atmosphere and heating or cooling history. Route ranking considers competing phases and outcomes alongside the target, allowing experimental evidence to guide which preparation should be attempted next. The output joins the route, its supporting evidence and the variables that most affect its prospects.
Condition-aware route ranking and outcome-prediction models connected to precursor, reaction and attempted-process evidence.
- Inputs
- Candidate structure or material identity · Precursor catalog and reaction evidence · Particle, mixing, atmosphere and thermal history · Equipment constraints and prior attempt outcomes
- Outputs
- Ranked precursor and process routes · Expected target and competing outcomes · Route uncertainty and follow-up evidence priorities
Within the engineConnects SYMGÈNE crystal candidates and MIXPROC process proposals to CINÉTA kinetics and the experimental action selector.
Scientific area / 04
Follow the physics outward.
CETTE-MÉSOLIEN · CETTE-CHAMPOP · CETTE-CINÉTA
Microstructure learning, neural solution operators and hybrid reaction dynamics connect process history to effective properties and physical fields.
Microstructure encoders · neural operators · hybrid kinetic models
Neural network family
CETTE-MÉSOLIEN
Microstructure and scale transfer
MÉSOLIEN connects materials chemistry and preparation history to the microstructures that govern engineering response. It combines composition, phase fractions, porosity, microscopy and processing evidence to represent plausible microstructures and distributions of effective properties. Shared batch effects and covariance can propagate through the scale-transfer chain. Observation models and discrepancy terms keep the mapping responsive to what experiments can identify.
Conditional microstructure representations and learned mappings from composition, processing and observations to effective-parameter distributions.
- Inputs
- Composition and recipe basis · Precursor state and process history · Microscopy, scattering, porosity, phase and moisture evidence · Intrinsic phase properties and uncertainty
- Outputs
- Microstructure realizations or latent microstructural states · Effective tensors and constitutive parameters · Covariance and uncertainty across the scale mapping
Within the engineSupplies material parameters and microstructure distributions to CHAMPOP field models, CINÉTA process dynamics and product-specific mechanics or transport solvers.
Neural network family
CETTE-CHAMPOP
Field prediction with neural operators
CHAMPOP learns the map from a physical problem specification to spatial or time-dependent fields. Geometry, material coefficients, sources, initial conditions and boundary conditions enter the model together. Its operator formulation supports repeated thermal, moisture, flow or mechanical calculations within a defined problem family. Numerical solver references and conservation checks connect rapid field inference to the equations and conditions it represents.
Neural solution operators using Fourier, DeepONet, graph or geometry-aware formulations selected for mesh, boundary and temporal requirements.
- Inputs
- Geometry, mesh or supported spatial representation · Material and constitutive coefficient fields · Initial and boundary conditions · Sources, loads and time history
- Outputs
- Spatial or temporal physical fields · Integrated or local quantities of engineering interest
Within the engineUses MÉSOLIEN effective parameters and CINÉTA reaction terms to connect material design with thermal, transport, moisture and mechanical product analysis.
Neural network family
CETTE-CINÉTA
Constrained reaction and process dynamics
CINÉTA combines learned kinetic information with explicit chemical and engineering structure. Neural components estimate rate parameters, closure relations or residual corrections, while stoichiometry and balance equations connect those quantities to species evolution. The same family can support a well-mixed reactor, a reduced plug-flow model or a transport-coupled process when its assumptions match the apparatus. It links reaction pathways to time, temperature, composition and operating conditions.
Hybrid neural kinetic or differential-equation model embedded in stoichiometric balances and the relevant transport or reactor equations.
- Inputs
- Species inventory and stoichiometric reaction structure · Temperature, pressure and relevant activities or concentrations · Flow, heat-transfer and reactor conditions · Kinetic and process observations
- Outputs
- Reaction-rate or closure estimates · Time-dependent species and process-state predictions · Yield, selectivity and balance quantities for the defined system
Within the engineConnects synthesis, mineralization, binder chemistry, electrochemistry and waste conversion with transport and field models.
Scientific area / 05
Ask a sharper electronic question.
CETTE-SUPRAC
Condition-aware models prioritise superconducting-material candidates, with separate conventional and unconventional research paths.
Composition and structure encoders · conditioned property heads
Neural network family
CETTE-SUPRAC
Superconducting-material screening
SUPRAC organizes superconductivity discovery around composition, structure, phase and measurement conditions. Its prediction heads can rank candidates or estimate a defined transition-temperature target while retaining pressure, doping and the criterion used to identify the transition. Conventional phonon-mediated calculations and strongly correlated research follow their appropriate physical routes. This allows material screening to connect to synthesis and characterization while retaining separate material and device objectives.
Condition-aware composition or structure encoders with ranking, transition-temperature and uncertainty heads; optional surrogates for specified electronic or phonon quantities.
- Inputs
- Composition, doping and crystal or phase information · Pressure and relevant physical conditions · Transition criterion, source provenance and uncertainty · Appropriate electronic or phonon descriptors for the selected route
- Outputs
- Conditioned candidate ranking or transition-temperature prediction · Uncertainty and parameter sensitivity · Property-specific synthesis and characterization priorities
Within the engineConnects SYMGÈNE and SYNTHÉVAL to superconductivity-specific electronic, phonon and measurement workflows; supplies relevant material evidence to FLUXCIR.
Scientific area / 06
From circuit to device behaviour.
CETTE-FLUXCIR · CETTE-SYNDÉCODE
Distinct model families support Coltrane circuit design and temporal quantum-error decoding, connecting materials, fabrication variation, control and measurement.
Circuit graphs · physical simulation · temporal decoding models
Neural network family
CETTE-FLUXCIR
Superconducting-circuit design and response models
FLUXCIR connects Coltrane circuit topology, device geometry and fabrication context to superconducting-circuit behavior. It supports the exploration of junction, capacitance, inductance, coupler and readout choices while accounting for process variation and packaging. Circuit spectra and transition matrix elements are connected with electromagnetic participation and relevant noise or loss models. Independent circuit and field solvers provide the physical reference for candidate designs.
Circuit and layout proposal models coupled to parameterized response surrogates, circuit Hamiltonians and electromagnetic extraction.
- Inputs
- Circuit graph, layout and fabrication stack · Junction, capacitance and inductance parameters · Coupler, readout and package configuration · Control bias, boundary conditions and fabrication variation
- Outputs
- Circuit and layout proposals · Energy spectra, transition matrix elements and coupling · Field participation, loss estimates and fabrication sensitivity
Within the engineCombines appropriate SUPRAC material information with circuit and electromagnetic analysis and provides device/noise context for SYNDÉCODE.
Neural network family
CETTE-SYNDÉCODE
Temporal quantum-error decoding
SYNDÉCODE uses the history of quantum-error syndromes to infer the information needed for decoding. The model’s input retains the code, noise and measurement-cycle context so that correlated errors and temporal structure can be represented. Its system role includes both decoding quality and execution behavior: latency, throughput, memory and the cycle budget matter alongside logical-error performance. The family connects learned decoding to the operating constraints of quantum control.
Temporal neural decoding network that relates syndrome sequences and code context to logical-state or correction decisions.
- Inputs
- Time-ordered syndrome or detection-event records · Code structure and distance · Noise, leakage and measurement-cycle context
- Outputs
- Logical-state, error-class or correction decisions under the selected decoder interface · Associated decoding confidence where calibrated
Within the engineConsumes code and noise models tied to Coltrane device characterization and returns decisions to a defined classical control and decoding interface.
Scientific area / 07
Model the operating conditions.
CETTE-ÉLECTROFLUX
Hybrid electrochemical models connect fuel-cell materials with reactions, species transport, water management, heat flow and degradation.
Conditioned property models · reaction and transport coupling
Neural network family
CETTE-ÉLECTROFLUX
Fuel-cell electrochemistry, transport and degradation
ÉLECTROFLUX links fuel-cell materials to operation at the cell scale. The family relates catalyst, membrane, ionomer and porous-layer properties to electrochemical losses, water management, heat and gas transport. Geometry, humidity, pressure, temperature and load history remain part of the prediction. Its initial chemistry focuses on proton-exchange-membrane cells, allowing polarization and degradation analysis to remain tied to a defined electrochemical system.
Conditioned material and cell-response surrogates with calibrated electrochemical and transport closures; connected field and temporal model components.
- Inputs
- Membrane, ionomer, catalyst and porous-layer properties · Flow geometry, cell area and compression · Gas composition, pressure, flow and humidity · Temperature, conditioning and load history
- Outputs
- Polarization and resistance response · Water, heat and transport quantities · Degradation indicators, gas consumption and power at a declared system boundary
Within the engineConnects catalyst/interface information, MÉSOLIEN microstructure, CINÉTA kinetics and CHAMPOP transport to cell-level design and operation.
Scientific area / 08
Choose the reference carefully.
CETTE-ONDEQ · CETTE-EMBEDQ
Neural wavefunctions and quantum embedding connect electronic structure to targeted reference calculations, with separate Hamiltonian, solver and boundary specifications.
FermiNet/Psiformer-lineage wavefunctions · embedding · selected quantum algorithms
Neural network family
CETTE-ONDEQ
Variational neural electronic wavefunctions
ONDEQ represents an electronic wavefunction with a neural model and optimizes it through variational Monte Carlo for a specified Hamiltonian. The calculation retains the nuclei, electron count, charge and spin sector together with pseudopotential choices where applicable and boundary conditions. Variational energies are reported with sampling and systematic-error analysis. This provides a higher-cost electronic reference route for questions where improved treatment can change a material decision.
Antisymmetric variational neural wavefunction models in the FermiNet/Psiformer research lineage, including attention-based formulations.
- Inputs
- Hamiltonian and nuclear geometry · Electron count, charge and spin sector · Boundary conditions and pseudopotentials where applicable · Trial-state family and optimization configuration
- Outputs
- Variational energy · Sampling uncertainty and optimization evidence · Systematic-error assessment for the defined reference calculation
Within the engineSupplies targeted electronic reference evidence to the method graph and material-discovery workflows, with EMBEDQ as a separate optional embedding route.
Quantum computing workflow
CETTE-EMBEDQ
Quantum embedding and variational eigensolver research
EMBEDQ places a selected electronic subproblem inside an explicit embedding environment. The workflow defines orbitals, electron counts, bath construction and self-consistency before selecting an appropriate classical or variational quantum solution route. It connects the embedded calculation back to the larger material problem through an error budget covering both solver and embedding approximations. EMBEDQ is the engine’s quantum-computing workflow alongside the ONDEQ neural wavefunction family.
Quantum embedding and variational quantum eigensolver workflow with explicit active-space construction, bath treatment and classical reference comparison.
- Inputs
- Electronic Hamiltonian and active-space definition · Orbital and electron counts · Bath construction and embedding assumptions · Solver, circuit and computational-resource configuration
- Outputs
- Embedded electronic energies or supported observables · Convergence and self-consistency evidence · Solver and embedding error estimates
Within the engineProvides an optional specialized reference path in the same method-compatibility graph used by atomistic and electronic-structure workflows.
Scientific area / 09
Make computation count.
CETTE-INVAR · CETTE-COMPACT · CETTE-NEUROCIR
Invariant baselines and compact models establish useful reference points. A distinct accelerator-design program connects workloads, memory, mapping and circuit implementation.
Invariant models · distillation and compression · workload co-design
Neural network family
CETTE-INVAR
Invariant property learning and baseline inference
INVAR retains a direct route from a material representation to a defined property. The family supplies simpler invariant predictors that support practical inference and establish a comparison for more elaborate geometric architectures. Each property head uses its own label definition, units, normalization and conditioning variables. Pooling follows the physical quantity: a per-system extensive output and an intensive material property require different constructions.
Invariant structural or property-prediction network with observable-specific pooling and output heads.
- Inputs
- Composition or structural representation supported by the selected model · Requested property and required conditions
- Outputs
- Defined scalar or other supported invariant property · Associated domain and uncertainty information
Within the engineProvides a retained baseline for atomistic and property models and a compact route for property-specific inference.
Neural network family
CETTE-COMPACT
Compression and efficient deployment
COMPACT carries trained scientific capabilities into a smaller deployment model. Distillation and compression are evaluated on the outputs that matter to the application, including forces, stress or trajectories when those endpoints are used. The deployment profile records precision, memory and execution cost alongside scientific accuracy so that faster inference retains the behavior required by the task.
Compact or distilled neural networks with task-specific inference and derivative qualification.
- Inputs
- Inputs of the supported teacher or task model · Training or distillation data for the intended deployment domain · Runtime and numerical precision profile
- Outputs
- Task predictions in a compact inference package · Required energy derivatives when supported by the deployment endpoint
Within the engineSupports lower-cost execution of qualified model families and connects scientific accuracy requirements to NEUROCIR workload and hardware analysis.
Neural network family
CETTE-NEUROCIR
Neural accelerator and hardware/software co-design
NEUROCIR connects the structure of an AI workload to accelerator architecture, software mapping and physical implementation. It models tensor shapes, precision, data movement, processing elements, memory and interconnect together, so search can address practical bottlenecks. Candidate architectures and mappings move through analytical models, simulation and implementation checks. The family also links scientific inference workloads to accuracy, latency and energy objectives under a defined hardware and system boundary.
Workload and architecture representations with compact latency, energy and area surrogates; proposal or mapping models connected to implementation analysis.
- Inputs
- Workload graph, operators and tensor shapes · Precision, accumulation, batch and sparsity assumptions · Memory hierarchy, processing elements, interconnect and dataflow · Compiler mapping, process technology, clock and thermal limits
- Outputs
- Architecture and mapping proposals · Latency, energy and area estimates for the specified workload · Implementation and thermal tradeoffs within the selected analysis route
Within the engineUses the actual workloads of ÉQUIAT, COMPACT and other engine families to guide computing design; connects candidates to independent mapping, RTL and physical-design flows.
02 / Le laboratoire des idées
Change the
question.
A material is a starting point.
Explore how the property changes the research path.
Explore a property
The observable
How does heat move?
Thermal conductivity01 / Define the question
Conditions give it meaning.
Temperature · direction · phase · defects
An energy prediction alone does not establish thermal conductivity.
03 / Les programmes
New matter.
Real purpose.
From resilient infrastructure and resource recovery to energy conversion and computing, our programs connect materials research with the demands of an engineered product.
01Water & infrastructure
HydroGrid
Permeable paving systems
A recycled-polymer grid and engineered aggregate system designed to support paved surfaces while allowing stormwater to infiltrate and be stored within the pavement assembly. The program connects material behavior with the structural and hydraulic conditions of an installed system.
Inside the research
Development brings together polymer mechanics, grid–aggregate contact, load distribution, drainage, and aging. Evaluation considers rutting, repeated loading, and water movement under defined installation and site conditions.
Discuss this programDevelopment focusStructural support, stormwater management & service life
02Construction materials
Carbon-storing concrete
Composition, processing, and performance
Cementitious formulations being developed to combine carbon dioxide uptake with the mechanical and durability requirements of construction. Our approach connects formulation and curing choices with phase development, microstructure, and material performance.
Inside the research
The research program evaluates carbon uptake alongside strength, transport properties, and long-term exposure. Uptake measurements and lifecycle carbon accounting address different questions; a carbon-negative designation requires a defined, verified accounting boundary.
Discuss this programDevelopment focusCarbon uptake, mechanical performance & durability
03Resource recovery
Waste-to-fuel systems
Integrated conversion-process concepts
Process concepts for converting selected waste plastics, biomass, and agricultural residues into useful liquid fuels. AI-assisted process development examines how feedstock composition, conversion conditions, and catalyst behavior influence product quality and resource use.
Inside the research
Reaction and transport models guide experiments, while material and energy balances connect individual processing steps. Validation includes product analysis, conversion and selectivity measurements, and the effects of catalyst aging.
Discuss this programDevelopment focusFeedstock conversion, product quality & process efficiency
04Carbon management
Sorbents & reactors
Capture materials connected to process design
Materials and reactor research for carbon dioxide capture, linking sorbent properties to the conditions in which a process must operate. The platform supports screening of adsorption behavior, kinetics, and transport before committing to more costly experiments.
Inside the research
Evaluation examines uptake and breakthrough, selectivity, cycling, and regeneration requirements where applicable. Reactor and process models help relate material-level results to flow, energy demand, and system-level performance.
Discuss this programDevelopment focusCapture behavior, material stability & reactor performance
05Quantum technology
Coltrane
Superconducting processor research
A fluxonium-based quantum processor research program connecting circuit design, materials, fabrication variation, and device characterization. A separate sensing track explores related measurement applications.
Inside the research
CETTE-FLUXCIR supports circuit and device modeling; CETTE-SYNDÉCODE addresses temporal quantum-error decoding. The validation path links fabricated-device measurements, calibrated control and readout, leakage, and logical performance.
Discuss this programDevelopment focusCircuit design, device characterization & error decoding
06Hydrogen & energy
Hydrogen fuel cells
From functional materials to cell behavior
Research connecting catalyst and membrane selection with electrochemical reactions, species transport, water management, and heat flow. CETTE-ÉLECTROFLUX supports the transition from material screening to component and cell models.
Inside the research
Assessment retains operating chemistry, temperature, pressure, humidification, and load. Cell measurements and degradation tests establish performance; net system efficiency also accounts for hydrogen consumption and supporting equipment.
Discuss this programDevelopment focusElectrochemical performance, transport & durability
07Advanced electronic materials
Superconductors
Screening under defined physical conditions
A materials research program that prioritizes candidate superconductors using material-family and pressure-dependent information, with selected electronic-structure and phonon calculations where scientifically appropriate.
Inside the research
CETTE-SUPRAC supports candidate screening. Reproducible synthesis, phase identification, electrical transport, and magnetic measurements provide the experimental route to evaluating superconductivity and useful engineering properties.
Discuss this programDevelopment focusCandidate selection, physical characterization & engineering use
08AI hardware
Neural processing units
Workload-aware accelerator design
Accelerator research that connects neural-network workloads with compute architecture, memory hierarchy, data movement, and physical implementation. CETTE-NEUROCIR supports architecture exploration, workload mapping, and circuit-design modeling.
Inside the research
Comparisons retain precision, bandwidth, technology, and clock assumptions. The validation path progresses through simulation, logic verification, physical design, and measured workloads to assess useful throughput and system energy.
Discuss this programDevelopment focusWorkload performance, memory efficiency & physical implementation
Our portfolio spans software, materials, and device development. Performance is evaluated under defined test conditions. Contact us to discuss current program status, validation, and partnership opportunities.
04 / La méthode
Predict.
Verify. Go further.
Scientific ambition needs a record that can be challenged. Source identity, method compatibility, uncertainty and independent verification belong inside the research loop.
Define the question.
Specify the material or device, the property of interest, operating conditions, and practical constraints. Preserve the source and meaning of each observation.
Choose the next test.
Combine models and scientific judgment to select useful simulations and experiments. Compare candidates against reference methods, uncertainty, and the cost of verification.
Learn from the evidence.
Record measured results, deviations, and unsuccessful tests. Evaluate model updates against independent observations before using them to guide the next campaign.
05 / Parlons recherche
Let’s build
what’s next.
Our office223 Bedford Ave.Brooklyn, NY 11211Start a conversation about research, technology, or a potential partnership.
General inquiriesinfo@cette.aiOffice(347) 716-7434Please keep initial inquiries non-confidential.