Leveraging Agentic AI — Modernizing Visual Simulation Workflows in HYSYS + Aspen Plus
_____________________________________________________________________________________
Case Study Snapshot:
Aspen's simulators (HYSYS and Aspen Plus) are built on powerful visual chemical engineering workflows, but the legacy experience was deterministic and manual—too many steps, constant tuning, error-prone decisions, and "hidden levers" that required expert knowledge. I pushed the platform into a new generation by embedding agentic AI directly into the simulation experience, starting with a core workflow: creating a feed for a distillation column. The approach turned expert behavior into a guided, explainable system—moving complexity out of the user's cognitive load and into the platform, with a mixed autonomy model that assists without taking control.
The AI layer was deeply integrated into the visual workspace: inline tooltips, guided paths replacing tribal knowledge, and a "View Agent" entry point for deeper reasoning. Agent capabilities included recommendations, automated multi-step changes with confirmation, scenario generation, and root-cause diagnosis—all designed with explicit trust guardrails: explainability via citations and confidence signals, bounded action space, audit trails, and confirm-before-change gates. Results from a year of usability studies and customer pilots showed 120% improvement in completion speed, 95% task success rate, and zero high-severity failures—modernizing Aspen's core advantage while maintaining trust, control, and auditability.
_____________________________________________________________________________________
Context:
Aspen’s simulators (HYSYS and Aspen Plus) are defined by powerful visual chemical engineering workflows. I pushed that foundation into a new generation by embedding agentic AI directly into the simulation experience—starting with a core workflow: creating a feed for a distillation column. Primary users included process engineers, operators, planners, reliability teams, and chemists/chemical engineers.
Problem:
The legacy flow was deterministic and manual: too many steps, constant tuning, error-prone decisions, slow iteration, and “hidden levers” that required expert knowledge. Users routinely got stuck and escalated to senior engineers for next steps. Outcome quality varied with experience level and the workflow took significantly longer than it needed to.
Approach:
We turned expert behavior into a guided, explainable system—moving complexity out of the user’s cognitive load and into the platform. The AI layer was built to assist without taking control, focusing on speed, accuracy, and confidence.
Key moves:
Embed assistance inside the workflow, not in a detached chat UI
Design a mixed autonomy model: suggest-only and execute-with-confirmation
Make reasoning visible so users could trust and verify
Keep the visual workspace clean by shifting deterministic complexity “behind the scenes”
Agentic capabilities embedded into the simulator experience:
Core agent behaviors
Recommendations and next-best actions
Automated multi-step changes (with confirmation)
Monitoring + alerts
Scenario generation
Root-cause diagnosis when outcomes deviated or convergence issues appeared
Agent action space
Change parameters
Insert equipment
Tune constraints
Run simulations
Compare scenarios
UX integration:
Inline tooltips for context-sensitive guidance
A dedicated “View Agent” entry point for deeper reasoning and configuration
Guided paths users could follow end-to-end, replacing tribal knowledge and screen-hopping
We designed explicitly for trust, safety, and auditability:
Explainability via citations, model trace, sensitivity analysis, and confidence signals
Permissions, bounded action space, and approval gates
Audit trail and rollback for reversibility
Execution constrained to confirm-before-change for consequential actions
Outcomes
Measured through usability studies, A/B tests, and customer pilots over a one-year period:
120% improvement in completion speed
95% task success rate
<5 iterations to converge on average
Zero high-severity failures in the measured scenarios
Why it mattered:
This didn’t bolt AI onto a legacy product—it modernized Aspen’s core advantage: visual engineering workflows. By embedding agentic assistance directly into the simulation flow, we made expert-level execution more accessible, reduced friction on complex tasks, and increased both speed and outcome quality—without sacrificing trust, control, or auditability.