Conjugate Heat Transfer Optimization for Turbine Blade Thermal Performance Using Field-Driven Design
Interview with Max Gaedtke + Markus Lempke of nTop + Siemens Energy
nTop and Siemens Energy will present a closed-loop parametric optimization workflow for turbine blade internal cooling at CDFAM Barcelona, demonstrating what becomes possible when implicit geometry and high-fidelity simulation are unified in a single design environment.
The interview below covers the technical architecture behind this workflow and the collaboration between nTop and Siemens Energy that connects computational design research to industrial turbo-machinery requirements.
Can you start by describing nTop’s role in this collaboration with Siemens Energy, and give us an overview of what you’ll be presenting at CDFAM?
At CDFAM, we’ll demonstrate a closed-loop parametric optimization on a turbine blade internal pin fin array configuration.
The engineering challenge here is fundamental: gas turbine efficiency improves with higher inlet temperatures, but these temperatures routinely exceed material limits. Internal cooling channels circulate bleed air through the blade to keep metal temperatures survivable—but every cooling passage and feature costs aerodynamic efficiency through pressure drop and reduced mass flow to the main combustion path. The design objective is maximizing heat transfer effectiveness while minimizing this penalty, within geometric constraints that push the limits of what conventional CAD and meshing tools can reliably handle. Compounding this, the design space has grown exponentially in recent years—from evaluating hundreds of variants to thousands, and now pushing toward tens of thousands with surrogate modeling and AI design agents. Traditional workflows simply cannot scale to meet this demand.
This is where nTop’s architecture becomes essential. Complex internal cooling geometries—serpentine passages, pin fin arrays, rib turbulators—frequently break traditional B-rep workflows at exactly the aggressive configurations that yield performance gains. Our implicit geometry engine guarantees robustness across the entire parameter space, while GPU-native CHT simulation makes iterative optimization practical rather than theoretical.
The key message: by eliminating the traditional handoff between CAD, meshing, and simulation, we can run high-fidelity CHT as an inner-loop optimization objective rather than a final verification step. We’ll show concrete results on pin positioning optimization, balancing thermal performance against pressure drop.
nTop provides the integrated design-and-simulation environment that makes this workflow possible. Our field-driven geometry engine generates and modifies complex internal cooling structures, while our GPU-native conjugate heat transfer solver evaluates thermal performance—all within a single platform. Siemens Energy brings deep turbomachinery expertise and defines the industrially relevant design constraints and performance targets.
Your framework integrates field-driven design with Lattice Boltzmann-based conjugate heat transfer simulation. What specific advantages does this architecture provide over conventional design-simulate workflows in turbomachinery applications?
Three fundamental advantages emerge from this architecture.
First, geometric robustness: nTop’s implicit representation guarantees watertight, manufacturable geometry for every parametric variant—no failed meshes, no manual repair. In conventional workflows, aggressive design changes frequently break the CAD-to-mesh pipeline.
Second, evaluation speed: our GPU-native LBM solver achieves roughly 200x reduction in time-to-solution compared to traditional finite-volume methods. A simulation that previously took hours now completes in under two minutes on consumer hardware.
Third, architectural unity: because both geometry and physics operate on the same implicit field representation, there’s no mesh regeneration between design iterations. This eliminates what is often the dominant time cost in optimization loops.
Together, these enable exploration of design spaces that would be computationally prohibitive with conventional tools—or simply impossible due to lack of geometric robustness.
How does nTop’s implicit geometry engine handle parametric changes in turbine blade cooling structures without requiring mesh regeneration, and how does this impact the speed and reliability of the optimization process?
nTop represents geometry as continuous signed distance fields rather than B-rep surfaces. When a parameter changes—pin position, channel dimensions, lattice density—the field updates analytically. The solver then voxelizes this field directly onto the computational grid.
This has two consequences for optimization. Speed: we avoid the surface mesh and volume mesh generation steps entirely—the solver voxelizes directly from the implicit field, which in complex internal geometries eliminates what often dominates total iteration time. Reliability: the implicit representation is inherently watertight and self-consistent. There are no degeneracies, no sliver elements, no Boolean failures. Every point in parameter space produces a valid geometry that the solver can evaluate.
For turbine blade cooling specifically, where internal passages involve tight radii, thin walls, and complex intersections, this robustness is essential. Conventional parametric CAD frequently fails at exactly the aggressive design configurations that optimization algorithms want to explore.
Can you explain how the GPU-native solver manages conjugate heat transfer without explicit fluid-solid interface modeling, and what types of data are exchanged between the design and simulation components during optimization?
The solver uses a Lattice Boltzmann Method formulation where thermal transport is governed by a total enthalpy approach. Fluid and solid regions are distinguished by their material properties—thermal conductivity, heat capacity, density—but the transport equations remain continuous across the interface. There’s no explicit interface tracking, no coupled boundary conditions to enforce. The physics simply emerges from the property fields.
During optimization, the design component passes the implicit geometry field and material assignments to the solver. The solver returns scalar objectives—peak temperature, average surface heat flux, pressure drop—and optionally full field data for visualization. The optimizer then proposes new parameter values based on these objectives, and the loop repeats.
Because both geometry and physics are field-based, this exchange is lightweight. We’re passing continuous functions, not tessellated meshes.
You mention validation against finite-volume baselines. What were the criteria for this validation, and how generalizable are the results to other thermal management problems outside of turbine blade cooling?
We rigorously validated specifically against the 3-Fin Heat Sink benchmark, comparing nTop Fluids against OpenFOAM’s chtMultiRegionFoam solver. The primary criteria were peak chip temperature and channel pressure drop—the quantities that matter for thermal design decisions.
At matched resolution, we achieved peak temperature agreement within 0.5% and pressure drop within 5% of the finite-volume baseline. Importantly, we also characterized the behavior at coarser resolutions: at 2x grid spacing, temperature over-prediction reaches roughly 20%, but physical trends remain correct. This gives users a quantified accuracy-speed tradeoff for early design exploration versus final verification.
It’s worth noting that conjugate heat transfer is currently a public beta feature in nTop Fluids, and solver validation is never truly finished—it’s an ongoing process. We execute validation continuously through our CI pipeline, expanding benchmark coverage as we encounter new application domains and edge cases.
Regarding generalizability: the underlying physics—forced convection with conduction through solid structures—applies broadly. Heat exchangers, cold plates, electronics cooling, battery thermal management—all share the same governing equations. The validation transfers directly. What changes between applications are the relevant Reynolds numbers, conjugate ratios, and geometric complexity, but the solver formulation accommodates these without modification.
What do you hope to share with and learn from other participants at CDFAM, especially with regard to industrial adoption of automated, high-fidelity optimization workflows?
We want to demonstrate that high-fidelity thermal simulation can be fast enough to sit inside optimization loops—that this is no longer a future capability but something achievable today on standard workstation hardware. The implications for DfAM are significant: rather than designing for manufacturability and then verifying performance, teams can optimize for performance within manufacturing constraints simultaneously.
What I’m keen to learn is where participants see potential applications for this technology stack. Which thermal management challenges in their workflows would benefit from implicit modeling, rapid CHT evaluation, or automated parameter optimization? Are teams already thinking about extending this toward multidisciplinary design optimization—coupling thermal performance with structural or manufacturing objectives? CDFAM tends to surface these application-driven conversations more clearly than any survey could.
I’m also curious how other organizations are thinking about the role of AI surrogates versus high-fidelity solvers. Our benchmarks show the GPU-native approach competitive with or faster than PINN-based approximations, which raises interesting questions about where surrogate models add value and where direct simulation is simply sufficient
The work presented here reflects a broader shift underway in engineering design: high-fidelity simulation moving from verification to optimization, and computational geometry becoming a first-class engineering tool rather than a downstream task.
These are precisely the developments CDFAM exists to track and advance. If the questions raised in this interview, around solver architecture, surrogate models versus direct simulation, and multidisciplinary optimization, are relevant to your work, the conversations happening at CDFAM Barcelona will be worth your time.
CDFAM brings together practitioners and researchers working at the leading edge of computational design, AI, and machine learning across engineering, architecture, and software development. Attending is the most direct way to engage with the people driving this field forward, including the nTop and Siemens Energy team presenting this work. Registration now to join the conversation.
