Accelerate Simulations & ML Surrogates

Success stories

Redefining Basin Simulation with HPC and Machine Learning

AI/ML

Desktop GUI

HPC

Scientific Computing

Basin cross-section · schematic

Geologic time150 Ma

Burial
Maturation
Migration
Accumulation

Burial, maturation and charge through geologic time.

Source rock
Carrier bed
Seal
Oil window
Hydrocarbons

About the Task

Petroleum-systems simulators carry decades of validated physics, and many of them still run as batch jobs held together by scripts. Here is how our scientific computing team modernizes this kind of software: we tune the numerical core for current clusters, build a desktop suite around it, and use the faster solver to train ML models that make large uncertainty studies affordable.

results

Faster simulations

results

Massive overnight modeling

Services used

Build Strategy

5–10×Shorter simulation runs on reference models

≥75%Strong-scaling parallel efficiency at 1,024 cores

≥1,000×Faster scenario evaluation with ML surrogates

10,000+Scenarios in an overnight sensitivity study

Design targets we engineer and test against. Real gains depend on the code base and the hardware, so we measure a baseline before committing to numbers.

The table of content

Overview

‍Context

Why basin simulators are hard to modernize

‍

A basin model reconstructs how a sedimentary basin was buried, heated and charged with oil and gas over tens to hundreds of millions of years. It couples burial history, heat flow, source-rock kinetics, pore pressure and multiphase migration on 3D grids that change shape through geologic time.

The numerical cores behind these models are usually a mix of Fortran and C++ that has been checked against real basins for years. Nobody wants to rewrite that physics, and nobody should. The work is in everything around it: parallel performance, input and output, set-up tools, visualization, and a way to run thousands of cases instead of a handful.

It takes three kinds of engineers who rarely sit in one team: HPC developers who are comfortable in old Fortran, desktop developers who know Qt and OpenGL, and ML engineers who understand what a surrogate can and cannot replace.

‍

Challenge

What usually slows modeling teams down

‍

A typical set-up: the simulator and the tools around it are separate applications connected by file exports.

RuntimeCode written for earlier hardware leaves most cores and GPUs idle, and serial output turns into the bottleneck as grids get finer.
Too many hand-offsData preparation, parameter set-up, uncertainty runs and 3D review happen in different tools, joined by exports and format conversions.
Slow model set-upGetting horizons, wells, lithologies, kinetics and boundary conditions into a valid model is mostly manual work.
Few scenariosIf one full-physics run takes hours, a sensitivity study stops at a few dozen cases.
Knowledge in a few headsA handful of senior modelers know every step of the workflow, and their time is the scarcest resource on the team.

‍

Architecture

The solver stays at the center

‍

We don't replace a validated simulator. We make it faster and give every other component one versioned API to reach it. The desktop suite, Python notebooks and ML agents all start the same runs and read the same results.

Desktop suite

Qt 6 and OpenGL. One project workspace from raw data to final maps.

Project Manager
Model Builder
Run Monitor
Uncertainty Studio
4D Viewer

ML & agents

Surrogates and assistants that call the same API a person would.

Set-up assistant
Run watchdog
Sensitivity assistant
Surrogate service · ONNX
On-prem LLM

Application framework

C++ core with a Python API through pybind11.

Plug-in SDK
Project & data model
Job orchestration
Provenance & audit
Tool adapters

Data layer

Open, parallel formats with metadata and lineage.

HDF5 / XDMF store
RESQML exchange
LAS · SEG-Y · ZMAP importers
Metadata catalog

Compute · Linux cluster

The existing simulator, profiled, optimized and containerized.

Basin simulation engine
Fortran · C++
MPI + OpenMP
GPU kernels · CUDA / OpenACC
Slurm · Apptainer

Existing simulator, optimized Existing in-house tools, connected through adapters

Speed before screensWe start with the solver. Every tool built later benefits from the faster engine, and performance problems can't hide behind a UI.

One API for people and scriptsThe GUI, notebooks and agents go through the same API, so automation can't skip validation or the audit log.

Nothing leaves the buildingThe suite, the ML models and any language model run on the client's own infrastructure.

‍

HPC Engineering

Making the numerical core scale

‍

Strong-scaling acceptance envelope

Speed-up relative to 16 cores. We agree an efficiency floor up front and check every benchmark run against it.

01
Profile before touching codeVTune, Advisor, Nsight Systems and Score-P give a hot-spot map on reference models and a baseline everyone agrees on.
02
Modernize the hot loopsFortran 2018 and C++20 refactoring, vectorization and cache-friendly data layouts where the profile says it matters.
03
Hybrid parallelismMPI domain decomposition across nodes, OpenMP threads inside a node, and load balancing for irregular basin grids.
04
GPU kernels where they pay offCUDA or OpenACC for kinetics, PVT/flash and linear solvers, but only for kernels that dominate the runtime.
05
Parallel I/OHDF5/XDMF with chunking and compression, checkpoint/restart, and RESQML exchange with interpretation tools.
06
Guard the physicsA golden-case regression suite with agreed numerical tolerances runs in CI on the cluster for every change.

‍

ML & Agents

Where machine learning helps, and where it shouldn't decide

‍

Set-up assistantReads the data inventory, drafts a model set-up (stratigraphy, lithologies, kinetics, boundary conditions) and points out gaps and conflicts.
Run watchdogFollows convergence, timestep control and mass balance on running jobs, explains why a run failed and proposes a restart.
Sensitivity assistantSets up experimental designs, launches ensembles and shows which parameters drive temperature, maturity and charge.

Assistants work through the framework's Python API. Every action is logged, and anything that changes a model or a result waits for a person to approve it.

‍

Screening 10,000 scenarios

without 10,000 full runs

‍

Surrogates learn from ensembles of full-physics runs on the cluster. Gaussian processes and gradient boosting handle scalar outputs; neural operators and graph networks handle 3D fields such as temperature, vitrinite reflectance (%Ro) and pore pressure.

Each prediction comes with an uncertainty estimate. When it is too wide, the case goes back to the full simulator, and the result becomes new training data. Accuracy improves where the team actually works.

10,000scenarios

screened by the surrogate in minutes

200candidates

ranked by charge risk and uncertainty

20full runs

checked with full physics on the cluster

1view

P10 / P50 / P90 ready for review

‍

Delivery

Four workstreams, one code base

‍

The streams run in parallel and meet at shared integration releases, so the desktop and ML teams always build on the current solver instead of waiting for it.

ASolver & HPC

Led by an HPC architect

Baseline & profiling
Hybrid MPI / OpenMP
GPU kernels
Parallel I/O

BDesktop platform

Led by a Qt / OpenGL architect

Suite shell
Model Builder
4D Viewer
Plug-in SDK

CData & integration

Led by a solution architect

Importers
HDF5 project store
Tool adapters
Job orchestration

DML & agents

Led by an ML lead

Training ensembles
Surrogates
Assistants
MLOps

‍

Where each skill is used

Core skill for the stream Significant Supporting

A typical team at full speed: 24 specialists

‍

Targets

Design targets and how we verify them

‍

Metric	Baseline	Target	Verified by
Simulation wall-clock, reference models	Measured at kick-off	5–10× faster	Benchmark suite on the client's cluster
Strong-scaling efficiency	Measured at kick-off	≥75% at 1,024 cores	Scaling runs from 16 to 1,024 cores
Checkpoint and output I/O	Measured at kick-off	3× faster, half the storage	Parallel HDF5 with compression, same fields
Time to set up a reference model	Time study with modelers	−60%	Before/after walkthrough with the same team
Tools in the daily workflow	Several separate applications	One suite	Workflow walkthrough and user survey
Sensitivity study throughput	Dozens of full-physics runs	10,000+ scenarios overnight	Ensemble logs, surrogate plus full-physics checks
Surrogate accuracy on key outputs	n/a	≤5% relative error	Hold-out cases with calibrated uncertainty

‍

Risks

Engineering risks and how we handle them

‍

Numerical drift during optimizationHigh
Golden cases with agreed tolerances, bit-for-bit checks where feasible, and sign-off by the client's geoscientists before each release.
IP and data protectionHigh
We develop and deploy inside the client's environment, with role-based access and on-prem language models. Data stays where it is.
Undocumented tool interfacesMedium
Interface discovery early on, a versioned adapter layer with contract tests, and no changes inside the tools themselves.
Over-confident surrogatesMedium
Uncertainty estimates, physics checks, and an automatic fallback to the full simulator outside the training range.
Cluster differencesMedium
Spack and CMake builds, Apptainer containers, and CI runners on the target cluster from the first month.
AdoptionMedium
Modelers join sprint reviews, UX research happens with real models, and each release ships with training.

‍

Working with us

Why teams bring this work to Infinity

‍

100+ AI experts in 13 countriesResearch scientists, ML engineers and technology leads, combined with HPC and desktop engineers into one team for one code base.

InfinitySDLCOur AI-agent-driven delivery model covers discovery, architecture, environments, release orchestration and security on large code bases.

Standards-aligned deliveryISO/IEC/IEEE 12207 life cycle, ISO/IEC/IEEE 29119 testing, ISO/IEC 5338 and 42005 for AI systems, OWASP SAMM for software assurance.

Enterprise track recordBrands we have worked with include Shell, Samsung, IBM and BNP Paribas.

How a project usually starts

WEEKS 1–2Access and baselineBuild the current code, run it on reference models, profile it.

WEEKS 3–4ArchitectureTarget architecture, adapter design for existing tools, data-format decisions.

WEEKS 5–6Plan and targetsMeasured baselines, agreed targets, and a team plan for the first release.

Have a simulator that deserves better tooling?

Tell us about your code base and your hardware. We'll come back with a baseline plan and the people to run it.

Talk to us

‍

Our success stories

02/10/2026

Redefining Basin Simulation with HPC and Machine Learning

Rebuilding basin simulation software for high-performance computing and machine learning to deliver faster, data-driven subsurface insights.

AI/ML

Desktop GUI

HPC

Scientific Computing

Enterprise Agent Mesh platform illustration for the Compliance Agent guide, with blue hexagonal agents and a central governance shield.

September 2026

Compliance Agent: Evidence Logistics, Control Testing, and the Mesh’s Control Plane

Engineering guide 12/12: build a Compliance Agent for reproducible control testing, sealed evidence, curated crosswalks and continuous mesh conformance.

Infinity Technologies

Enterprise Agent Mesh Engineering Guides

AI/ML

Shared Enterprise Agent Mesh illustration for the Legal Agent engineering guide.

September 2026

Legal Agent: The Playbook Is the Program

Engineering guide 11/12: build a Legal Agent where executable playbooks, matter access, contract lineage and privilege boundaries govern model-assisted review.

Infinity Technologies

Enterprise Agent Mesh Engineering Guides

AI/ML

Shared Enterprise Agent Mesh illustration for the Marketing Agent engineering guide.

September 2026

Marketing Agent: Generation Is the Commodity, the Constraint System Is the Product

Engineering guide 10/12: make generation subordinate to market-scoped claims, rights, channel rules, attribution discipline and measurable brand controls.

Infinity Technologies

Enterprise Agent Mesh Engineering Guides

AI/ML

Shared Enterprise Agent Mesh illustration for the Sales Agent engineering guide.

September 2026

Sales Agent: Make the CRM True, Then Worry About Selling

Engineering guide 09/12: make CRM state evidence-backed before generating selling assistance, with governed field updates, customer commitments and cross-agent handoffs.

Infinity Technologies

Enterprise Agent Mesh Engineering Guides

AI/ML

Shared Enterprise Agent Mesh illustration for the Customer Support Agent engineering guide.

September 2026

Customer Support Agent: Containment Quality, Not Deflection Rate

Engineering guide 08/12: containment quality, account-scoped retrieval, guarded customer sends, escalation packets and evidence-driven support autonomy.

Infinity Technologies

Enterprise Agent Mesh Engineering Guides

AI/ML

Shared Enterprise Agent Mesh illustration for the Operations Agent engineering guide.

September 2026

Operations Agent: Durable Execution, Process Conformance, and the Connective Tissue of the Mesh

Engineering guide 07/12: durable workflow state, bounded judgement, process conformance, runbook safety and chaos-tested recovery.

Infinity Technologies

Enterprise Agent Mesh Engineering Guides

AI/ML

Shared agent-platform illustration for the Finance Risk engineering guide: connected hexagons surrounding a governance core.

September 2026

Finance Risk Agent: Evidence Assembly, Adversarial Review, and Model Risk Management

Engineering guide 06/12: deterministic treasury calculations, facility-specific covenant definitions, adversarial challenge and human decision authority.

Infinity Technologies

Enterprise Agent Mesh Engineering Guides

AI/ML

Enterprise Agent Mesh platform illustration for the Finance Agent engineering guide.

September 2026

Finance Agent: Numbers From Tools, Never From the Model

A production engineering guide to a Finance AI Agent where every figure comes from deterministic tools and immutable fact packs, while the model is limited to grounded narrative and workflow orchestration.

Infinity Technologies

Enterprise Agent Mesh Engineering Guides

AI/ML

Governed enterprise agent platform with specialized nodes around a protected control core.

September 2026

Procurement Agent: Segregation of Duties Encoded in the Tool Layer

Engineering guide 04/12: build a Procurement Agent where approvals, supplier banking and payment authority are structurally outside the model’s tool and credential boundary.

Infinity Technologies

Enterprise Agent Mesh Engineering Guides

AI/ML

Shared enterprise agent platform illustration used for the Supply Chain Agent engineering guide.

September 2026

Supply Chain Agent: Exception Narratives Over an Optimiser You Already Own

Engineering guide 03/12: build a Supply Chain Agent that triages planning exceptions, explains shortage causality with provenance and delegates quantities to deterministic solvers.

Infinity Technologies

Enterprise Agent Mesh Engineering Guides

AI/ML

Governed enterprise agent platform connecting specialist agents around a protected control core.

September 2026

HR Agent: Effective-Dated, Jurisdiction-Scoped Policy Retrieval

Engineering guide 02/12: build an HR Agent that resolves employee context before retrieval, answers against effective-dated policy and routes sensitive cases.

Infinity Technologies

Enterprise Agent Mesh Engineering Guides

AI/ML

Governed enterprise agent mesh: connected hexagonal agents around a protected platform core.

September 2026

Recruitment Agent: Evidence Assembly Under a High-Risk Regulatory Regime

Engineering guide 01/12: build a Recruitment Agent that assembles requirement-linked evidence, preserves provenance and keeps candidate decisions with humans.

Infinity Technologies

Enterprise Agent Mesh Engineering Guides

AI/ML

September 2026

AI Model Router Agent — Engineering Guide

Engineering guide 12/12: route tasks across Claude, Codex and self-hosted models using data policy, capabilities, evaluation scores, cost and availability.

Infinity Technologies