AI-Assisted Refinery Scheduling

AVEVA

Designing AI Assisted Refinery Scheduling

Reframing optimisation software as a decision partner for oil refinery schedulers

AVEVA’s optimisation suite combines crude assay science, refinery modelling, and economic planning to help energy companies make high-value operational decisions.

Schedule AI introduced machine-learning into the most

time-critical layer of that stack: short-term refinery scheduling.

My role was to translate complex optimisation outputs into clear decision support, enabling schedulers to generate and refine feasible schedules while maintaining full operational control.

Overview

Role: Lead Product Designer

Team: ML specialists, engineers, PM & domain specialists

Focus: Interaction design, research, workflow modelling, UI design

Impact

50% reduction in scheduler training time

Increased confidence in model-generated schedules

Clear optimisation boundaries between users and algorithms

Contributed to 1–3¢ per barrel margin improvements

The problem
Oil refinery scheduling is complex, high-risk work

Refinery schedulers coordinate the flow of crude and products across dozens of interconnected units, tanks and pipelines.

Every schedule must satisfy multiple constraints:

Inventory limits
Unit capacity
Crude properties
Economic targets

Creating a viable plan often requires hours of manual iteration — adjusting activities, running simulations, and reconciling conflicts.

Legacy tools provided calculation, but little guidance

Existing scheduling tools performed the complex calculations required to simulate refinery operations. However, identifying better scheduling options still depended largely on the scheduler, manually exploring scenarios, reconciling constraints, and interpreting results across a dense interface.

Legacy refinery scheduling interface
Dense workflows requiring manual construction and iterative feasibility checks.

Technological advances
present opportunity

New optimisation techniques make quality, system generated schedules possible

Advances in optimisation techniques, machine learning, and computational power made it possible to simulate refinery operations at scale and generate feasible schedule proposals under complex operational and economic constraints.

This opened the door for refinery scheduling to move beyond manual construction and validation toward AI-assisted generation and exploration of viable plans.

Key Insight

AI must be a decision partner,
not an automation layer

Early research revealed strong scepticism towards automated scheduling. Schedulers were willing to use AI assistance, but only if they could clearly understand what the system changed and why.

Rather than replacing manual planning, the system needed to:

Generate proposals
Surface trade-offs
Allow users to refine schedule proposals
Maintain explicit user control over critical decisions

Understanding the Domain
Establishing a thorough, shared knowledge

Before designing AI interactions, I immersed myself in refinery scheduling and its constraints. I did this through:

Interviews with schedulers, customer support and product managers
Observation of reconciliation workflows in the desktop tool
Mapping refinery constraints (space, time, material properties)
Creating diagrams to validate understanding with domain experts
Reviewing industry tools and literature

This grounding ensured the AI interaction model respected real operational risk.

Research Impact

Established shared language between design and optimisation teams
Clarified where automation could assist and where control must remain human

Understanding the scheduler decision lifecycle
A map of the key stages in refinery scheduling and the decisions made at each step, translated into user needs that informed the design of the scheduling assistant

The Design Challenge

A fundamental shift in workflow

Moving to system-generated schedule proposals meant feasibility could be established from the outset. However, introducing system-generated plans fundamentally disrupted established scheduling workflows.

The assistant needed to integrate system-generated proposals into familiar scheduling behaviours, supporting existing decision processes while enabling new ways to explore and refine schedules.

Scheduler Workflow: From Feasibility Checks to Feasible Proposals
AI-assisted scheduling generates feasible plans upfront, shifting the scheduler’s role from fixing conflicts to evaluating and refining options

Designing the Interaction Model

Designing for Multiple Entry Points

Schedulers rarely start from a blank plan. In practice they enter the workflow in different states:

• Creating a new schedule
• Adjusting an existing schedule
• Resolving infeasible activities

I mapped the decision branches across these scenarios to ensure the assistant could be invoked at any stage of planning, rather than forcing schedulers into a new rigid workflow.

Balancing optimisation with control

Optimisation models perform best when they have freedom to move activities. But schedulers hold operational knowledge the system must respect.

The design challenge was enabling schedulers to define what the

system could change and what must remain fixed.

I introduced the Fix / Flex / Forget framework and created supporting interactive prototypes to test with users and communicate with engineers

Fix — activity must not change
Flex — AI may adjust within defined bounds
Forget — AI may fully re-optimise

From framework to product interaction

The interaction model translated into the product: timeline and activity controls communicate schedule constraints while the side-panel interface captures operational parameters and economic inputs.

Outcomes & Impact

Schedule AI shifted refinery scheduling from reactive feasibility checks to structured, AI-assisted decision-making. As lead designer, I shaped the interaction model for a highly reactive, constraint-heavy domain — translating optimisation logic into clear mental models and controllable workflows.

What shipped:

Scheduling assistant deployed with system-generated feasible schedules
The Fix / Flex / Forget interaction model for defining optimisation freedom
Multi-entry workflow integration with the existing scheduling tool
Input patterns translating operational constraints and economic drivers into optimisation parameters
Customer onboarding with Idemitsu refinery operations teams

Final shipped product: