Hardest Part

Modeling the non-linear thermodynamics of distillation columns accurately enough to recommend setpoint changes without violating safety margins or product quality specifications.

Min Viable Scope

Deliver an advisory-only optimization model for fuel gas consumption on a single crude distillation unit. Explicitly exclude closed-loop write access to the control system, complex catalytic reactors, and site-wide steam balancing.

Cold Start Problem

Refineries refuse to connect unproven software to their distributed control systems. Overcome this by starting in open-loop advisory mode using historical sensor data from historian databases to demonstrate theoretical energy savings offline.

Time To First Value

3 to 6 months of site-specific thermodynamic calibration and open-loop observation before operators trust the recommendations.

Data Moat Available

true

Technical Difficulty

High

Wedge

Target independent US Gulf Coast refineries operating aging distillation units with volatile utility costs. Win by proving immediate energy reductions on a single, high-consuming subsystem like a crude distillation unit heater. Expand by rolling out control algorithms to secondary units like fluid catalytic crackers and eventually across the operator's entire multi-site portfolio.

Timing

Deep reinforcement learning models now handle multi-variable, nonlinear thermodynamic optimization in real time without diverging. Concurrent sensor proliferation across heavy industry provides the necessary granular, high-frequency data density that was physically absent two years ago.

Why This ICP

Mid-sized independent refineries face tighter margins and stricter carbon emission penalties than state-owned supermajors. They act faster on operational expenditure reductions and have fewer layers of bureaucratic approvals for new operational technology.

Size Of Prize

Approximately 700 operable petroleum refineries and 1,000 large petrochemical complexes globally face immense energy costs. Capturing a fraction of these energy savings justifies an annual software and service spend of $500k per facility, creating an $850M ($500k × 1,700 facilities) annual addressable market.

Gap Narrative

Oil refineries operate complex distillation columns and catalytic crackers that consume massive amounts of thermal and electrical energy. Existing Advanced Process Control systems optimize for yield and throughput but fail to dynamically adjust energy consumption against fluctuating ambient conditions and feedstock variability. Plant managers lack a system that continuously optimizes the energy-to-yield ratio in real time.

Defensibility

The system builds defensibility through site-specific data compounding and workflow lock-in. As the algorithm learns the unique thermodynamic quirks of a specific refinery's equipment over thousands of cycles, its efficiency surpasses any generalized off-the-shelf alternative. Replacing the system requires restarting the learning curve and risking immediate energy spikes, creating high switching costs.

Why This Thesis

A Service-as-Software approach allows the vendor to absorb the complex model training and integration, delivering guaranteed energy reductions rather than an empty tool. Refineries buy the outcome of lower utility bills without needing to hire specialized machine learning engineers to build or maintain the control logic.

SAM

~$200M-$400M targeting North American and European tier-one and tier-two refineries

SOM

~$10M-$30M realistic 3-year capture focusing on early adopters seeking immediate emission compliance and margin expansion

TAM

~1,000 global petrochemical refineries and processing plants × ~$500k-$1M/yr enterprise software spend ≈ ~$500M-$1B

Growth Rate

~12-18%/yr, driven by regulatory pressures on carbon intensity and volatile energy feedstock costs

Paid Comparable Spend

~$300k-$800k/yr per facility spent on legacy process simulation licensing, external thermal engineering consultants, and manual data aggregation

Kill Thresholds

• Integration with the local plant historian system takes longer than 30 days
• Operators physically execute fewer than 20 percent of the recommended setpoint changes
• Backtesting on historical data yields under 2 percent theoretical energy savings
• Sales cycle to secure a paid 90-day pilot exceeds 120 days

Leading Metrics

• Days to complete initial plant topology mapping
• Optimization scenarios run per operator per week
• Percentage of recommended setpoint changes executed by operators
• Prediction variance between modeled and actual BTU consumption
• Time spent resolving data ingestion errors per week

What Proves Right

Plant operators run the thermodynamic solver during their daily shift planning, executing at least four optimization scenarios per week. The solver identifies a minimum 3 percent verifiable reduction in specific energy consumption during the pilot phase. Plant managers sign $250,000 annual recurring contracts based on the achieved thermal efficiency gains.

What Proves Wrong

Process engineers reject the recommendations because the solver ignores undocumented physical constraints or unit safety margins. Plant IT blocks the deployment outright due to internal policies preventing high-frequency historian data egress to external environments. The actual realized energy savings yield a payback period longer than twelve months, causing finance teams to veto the renewal.