Your CSTR was designed for 92% yield. After commissioning, it delivers 78%. The reaction kinetics are well-characterised: lab-scale calorimetry confirmed the rate constants, and the thermodynamics are textbook. The CFD model shows good macromixing: the impeller Reynolds number is 50,000, turnover time is 30 seconds, and there are no dead zones. Everything looks right on paper. The 14% yield gap is real, reproducible, and expensive: $8M/year in lost product and waste treatment for a medium-scale specialty chemicals plant.
The problem is micromixing. Your fast competitive reaction (Damkohler number > 10) completes before the feed is molecularly mixed with the bulk. Reactant A encounters a local excess of reactant B at the injection point, forming the undesired byproduct. By the time the fluid is homogeneous, the damage is done. Your CFD model does not resolve these sub-millimeter concentration gradients, so it does not predict the yield loss.
This is not a niche problem. Any reactor with fast reactions (t_reaction < t_mixing) is micromixing-limited. Pharmaceutical crystallization, polymerization, neutralization, precipitation, fast organic synthesis, and nanoparticle formation all fall into this category. The standard CFD approach (Reynolds-averaged Navier-Stokes with global kinetics) misses the physics that controls selectivity.
The Three Scales of Mixing
Mixing in a stirred reactor occurs at three distinct length scales, each governed by different physics. Macromixing is bulk circulation driven by the impeller. The characteristic length scale is the vessel diameter (1-10 m), and the characteristic time is the circulation time (10-60 s). CFD captures this well. Mesomixing is turbulent dispersion and engulfment at intermediate scales. The characteristic length is the turbulent integral scale (1-100 mm), and the time scale is the turbulent diffusion time. Large-eddy simulation (LES) captures this, but at high computational cost. Micromixing is molecular diffusion at the smallest scales, the Batchelor microscale (1-100 um). The characteristic time is the engulfment time (1-100 ms). This is where molecules actually encounter each other and react.
For slow reactions (t_reaction >> t_mixing), macromixing controls performance and CFD alone is sufficient. For fast reactions (t_reaction << t_mixing), micromixing controls selectivity and CFD alone is insufficient. Most industrially important selective reactions fall in between, where meso- and micromixing both matter.
Why Standard CFD Gets Selectivity Wrong
Standard CFD solves the Reynolds-averaged Navier-Stokes (RANS) equations for velocity and the transport equation for species concentration. The averaging process smooths out turbulent fluctuations, including the concentration fluctuations that determine local reaction rates. When you compute the reaction rate from the average concentration field, you get the wrong answer for any nonlinear kinetics.
Consider a second-order reaction: rate = k[A][B]. The average reaction rate is k<[A][B]>, which equals k<[A]><[B]> only if the concentrations are uncorrelated. In a real reactor near the feed point, [A] and [B] are strongly anticorrelated: where A is high, B is low, and vice versa. The actual average rate can be 50-80% lower than the rate computed from average concentrations. This is the segregation effect, and it directly explains yield losses in mixing-limited reactors.
The closure problem is fundamental: RANS gives you average concentrations, but you need the joint probability distribution of concentrations to compute average reaction rates correctly. Various turbulent mixing models (Engulfment, IEM, Beta-PDF) attempt to close this gap, but they require model parameters (mixing rate constants) that are uncertain and geometry-dependent. They are calibration-dependent approximations, not first-principles solutions.
Multi-Scale Simulation: Resolving Micromixing from Molecular Physics
Physics-first multi-scale simulation addresses the micromixing problem by coupling continuum-scale CFD with molecular-scale transport. The approach works in three stages.
First, CFD solves the macroscale flow field: velocity, pressure, turbulent kinetic energy, and dissipation rate throughout the reactor. This identifies the regions where micromixing matters: near feed points, in high-shear zones around the impeller, and in regions with high local Damkohler numbers.
Second, molecular dynamics or mesoscale simulations resolve the sub-grid concentration field in those critical regions. Instead of assuming a mixing model, the simulation directly computes how concentration striations thin, fold, and diffuse to molecular homogeneity. The molecular diffusion coefficient, computed from first principles, determines the rate of the final mixing step.
Third, the reaction is computed on the resolved concentration field, capturing the segregation effect exactly. The local selectivity emerges from the local concentration field without model parameters or closure approximations. The overall reactor selectivity is the integral of local selectivities over the entire reactor volume, weighted by flow rates.
This approach is computationally more expensive than RANS with global kinetics, but it solves the right equations. For a reactor where a 14% yield loss costs $8M/year, the computation cost is negligible.
Feed Point Design: Where Micromixing Starts
The most practical lever for improving micromixing is feed point design. Where and how you add reactants determines the initial concentration gradients and the time available for mixing before reaction. A single pipe discharging into the bulk creates large initial gradients and poor micromixing. A multi-point injection system distributes the feed across the high-turbulence zone near the impeller, reducing local concentration ratios and improving selectivity.
Simulation quantifies the trade-offs. For a given reaction system, the simulation predicts selectivity as a function of number of injection points (1, 4, 8, 16), injection velocity (controlling initial momentum and entrainment), injection location (impeller tip, below impeller, surface), and feed concentration (dilute feed reduces local gradients but increases volume). Each configuration runs in hours. Testing each in the pilot plant would take weeks.
A pharmaceutical company optimizing a reactive crystallization found that moving the anti-solvent feed from a single dip tube to a 4-port injection ring near the impeller tips improved crystal size uniformity from RSD 45% to RSD 18%, while increasing yield from 82% to 91%. The simulation predicted these improvements within 3% accuracy before any physical modification was made.
Reaction Kinetics from Molecular Simulation
The multi-scale approach also addresses the other side of the mixing-reaction coupling: the kinetics themselves. Traditional reactor modeling uses global kinetics, rate expressions fitted to lab-scale data at well-mixed conditions. But the actual local conditions in the reactor (concentration, temperature, solvent environment) may differ significantly from the lab conditions where the kinetics were measured.
Quantum mechanical calculations provide intrinsic rate constants from transition state theory, without the assumption of ideal dilute conditions. For reactions in concentrated solution (typical of industrial reactors), activity coefficients, solvent cage effects, and ionic strength corrections can shift rate constants by factors of 2-10x. Molecular simulation computes these corrections from the local molecular environment, not from infinite-dilution assumptions.
This matters most for competitive reaction systems where selectivity depends on the ratio of rate constants. If both rates shift by the same factor, selectivity is unchanged. But if the desired reaction is more sensitive to local environment than the byproduct reaction (common when they have different molecularity or charge), then the local conditions determine selectivity, and you need accurate local kinetics to predict it.
Scale-Up: From Pilot to Production
Reactor scale-up is where mixing problems become expensive. A reaction that works at 1 L fails at 1000 L because the mixing time scales with vessel size while the reaction time does not. The Damkohler number increases with scale, pushing the reactor from the kinetically-limited regime (where chemistry controls yield) to the mixing-limited regime (where fluid mechanics controls yield).
Traditional scale-up rules (constant power per volume, constant tip speed, constant Re) do not maintain constant micromixing conditions. Power per volume is typically reduced at larger scale (from 5-10 W/kg at pilot to 1-3 W/kg at production) for mechanical and economic reasons, which directly increases the micromixing time. The result is lower selectivity at production scale, the classic scale-up penalty.
Multi-scale simulation predicts the scale-up penalty before you build the large reactor. Run the simulation at pilot scale and validate against pilot data. Then run at production scale with the proposed geometry, impeller, and operating conditions. The simulation predicts the yield and selectivity at production scale, accounting for the changed mixing environment. If the predicted yield loss is unacceptable, you modify the design (more injection points, higher local turbulence, staged addition) in simulation before committing to fabrication.
A specialty chemicals manufacturer used this approach to scale a diazotization reaction from 50 L pilot to 5000 L production. The simulation predicted a 6% yield loss at production scale with the initial design (single feed, standard Rushton impeller). It identified that a 6-blade pitched blade turbine with 8-point injection at the impeller plane would maintain pilot-scale selectivity. The modified design was built and commissioned directly at production scale, achieving 90% yield vs. the 91% target. Without simulation, the company estimated it would have needed 3-4 pilot campaigns at $200K each to find an acceptable scale-up path.
Implementation: When to Use Multi-Scale Reactor Simulation
Not every reactor needs multi-scale simulation. The decision depends on the Damkohler number. If your reaction is slow (Da < 0.1), standard CFD with global kinetics is sufficient. If your reaction is fast (Da > 1), micromixing controls selectivity and multi-scale simulation provides significant value. In between (0.1 < Da < 1), both approaches may be adequate, and the choice depends on the economic stakes.
Indicators that micromixing is limiting your reactor:
- Yield lower than lab-scale kinetics predict
- Selectivity sensitive to impeller speed or feed rate
- Scale-up from pilot to production caused yield loss
- Feed point location affects product quality
- Reaction time < 1 second (competitive reactions, precipitation, crystallization)
If you see these symptoms, your reactor is mixing-limited and standard CFD will not find the solution. Multi-scale simulation resolves the physics that controls your yield and provides quantitative guidance for feed point design, impeller selection, and scale-up. Explore PetroSim for reactor and process simulation, or discuss your reactor mixing challenges with our engineering team.