Digital AI has spent decades optimizing clicks and conversations, yet the industrial systems that generate energy, materials, and power remain largely under-optimized. Applied Computing argues the biggest untapped dataset on Earth is operational sensor data from plants, grids, and refineries, and unlocking it requires a physics-informed foundation model built on the modern Databricks Platform. The stakes are enormous: a 1 percent efficiency gain in the physical economy implies roughly 350 billion dollars annually, 380 million tons of avoided emissions, and progress on energy access for hundreds of millions of people.
Industrial assets already emit a continuous language of temperatures, pressures, flows, and vibrations. Most of it goes unanalyzed. Orbital, Applied Computing's foundation model for energy operations, is designed to listen to that language, ground insights in physics, and deliver recommendations engineers can trust. The Databricks Platform demonstrates how unified data and AI infrastructure makes this kind of domain-specific foundation model practical at industrial scale.
Why industrial AI must be grounded in physics
Industrial facilities generate hundreds of thousands of sensor events per second across time series, chemistry, and context, all intertwined. Most global compute targets digital problems, while less than 2 percent is applied to physical systems, leaving an estimated 90 percent of operational data unused. Economic, safety, and climate outcomes hinge on interpreting these signals before real-world events unfold.
A generalized, internet-trained model is ill-suited to this domain. Orbital is purpose-built for energy operations and uses physics as a guardrail so answers remain defensible in safety-critical settings. The company reports 99 percent accurate inference on why something happened, what is likely next, and what operators should do.
What Orbital is made of
Orbital blends five elements into one system: a domain-tuned LLM for energy, deep time-series analytics for optimization and anomaly detection, a physics model that validates every recommendation, retrieval over unstructured plant documentation, and a validation layer that binds them together. The physics grounding reduces hallucinations and enforces consistency with process realities.
Engineers also need provenance, not just predictions. Retrieval and vector search over procedures, P&IDs, and process flow diagrams help explain why the model recommends actions, connecting numerical signals to operational context.
Unifying messy plant data on the lakehouse
Plant data spans streaming historians and SCADA, structured records like maintenance and finance, and unstructured drawings and procedures. A lakehouse approach consolidates all three, with Unity Catalog governing bronze, silver, and gold layers so AI can operate over one controlled substrate. Each customer runs in a dedicated workspace, and models trained for a customer remain in that customer's Unity Catalog.
Industrial ingestion rarely fits cloud-era connectors. Applied Computing built connectors for legacy protocols such as OPC DA and historians like OSIsoft PI, Yokogawa ExaQuantum, and Honeywell PHD. Spark declarative pipelines (open sourced in Spark 4.1) and Python custom data sources let the same ingestion run on premises or in the cloud, while structured streaming handles checkpointing, recovery, and backfill.
Scaling AI without owning the infrastructure
Training and serving run on the Databricks Platform so the team avoids standing up proprietary stacks. Databricks AI Runtime provides serverless GPUs, including A10s and H100s. Because training is bursty, pay-per-use capacity proved more cost effective than expected for a startup. MLflow evaluations and experiments help keep a non-deterministic LLM returning consistent, physics-checked answers, with all models versioned in Unity Catalog.
For production, Mosaic AI Serving supports low-latency inference. Vector Search and document intelligence speed retrieval over unstructured plant documents. Lakebase powers a responsive user interface so engineers can interact with data and recommendations without lag.
Turning sensor streams into decisions engineers can trust
Industrial AI will only cross from demos to daily operations when models speak the language of physics and context. Orbital's approach, fusing time series, a domain LLM, physics validation, and unified data on the lakehouse, sketches a practical path from raw sensor streams to reliable, auditable decisions at scale.