Accelerated computing
Physical AI, at the plant edge
A fertilizer complex is remote, continuous and often restricted from cloud links, so perception has to be local and fast. Nitreum runs GPU inference at the plant edge and keeps training, simulation and optimization in the cloud, in a private cloud or on-prem. Items marked planned are on the roadmap, not running.
Jetson Orin / Thor · DeepStream · Metropolis · TensorRT · Holoscan
Edge perception
8 to 40 camera feeds per unit through DeepStream pipelines and TensorRT detectors and segmenters for granule size and shape, dust, caking, broken prills and bag defects; Holoscan fuses vision with analyzer, historian, emissions, acoustic and thermal streams at 1 to 10 Hz. Design target: 50 to 120 ms p95 inference.
Triton · NIM · NVIDIA AI Enterprise
Model serving
Vision, time-series, RAG and process-reasoning models behind one orchestrator, with canary deployment and per-site routing. NVIDIA AI Enterprise is the planned runtime for regulated deployments.
DGX / HGX · NeMo · Retriever · Guardrails
Training and reasoning (planned)
Granule vision, time-series prognostics and NeMo process-reasoning models fine-tuned from labeled images, process histories, engineer corrections and event logs. Design target: 2 to 5 model families and 500 to 2,000 GPU-hours per cohort.
Omniverse · Modulus · OVX
Digital twin (planned)
The Nitrotwin complex twin: physics-informed surrogates for the ammonia loop, evaporation, granulation, blending and emissions, with GPU-accelerated process and CFD sweeps run before a plant change.
Cosmos · Omniverse Replicator
Synthetic data (planned)
Rare dust, caking, wet-granule, broken-prill, lighting, belt-contamination, emissions-spike and upset scenarios generated and labeled as synthetic, validated against real plant events before a model learns from them.
cuOpt
Optimization
Throughput, recycle balancing, energy and emissions allocation and grade-changeover sequencing under constraints, recomputed on the GPU when the complex is disrupted.
RAPIDS
Data engineering
ETL over historian, quality, emissions and vision data for feature generation, quality correlation and cross-site benchmarks on the GPU.
Isaac · Jetson
Robotic handling
Bag palletizing, bulk loading, materials handling and inspection rovers, with perception and motion policies built and validated in simulation before they move a pallet.