Hazard Assessment via CFD


DBS, 31.03.16

1. Purpose

CHAM's Consultancy Team is frequently called upon to give advice on the possible consequences, and their relative probability, of hazards such as oil spills, air pollution, fires in car-parks, and the like.

The purpose of the present document is to set out, and to illustrate, the principles to be observed. Those principles include:

  1. Since hazard assessments always involve uncertainties about the what, where and how-much of the initiating event, simulation of a large number of scenarios is essential, in order that a worst-case scenario can be given special attention.

  2. Because CFD simulations of turbulent flows are never more than plausible approximations, and the presence of turbulence adds further uncertainty, quantitative predictions which are made must always be presented as being reliable only between estimated +/- % limits.

  3. When smoke-production and consequent visibility-impairment are in question, the error-band width must be especially wide.

  4. Although the numerical accuracy of CFD simulations can in principle be improved by the use of very fine grids, it is counter-productive to incur their associated expense only to achieve gains which are negligible compared with the uncertainties just enumerated.

  5. To summarize, the best analysis is a well-balanced one in which the amount of attention paid to an individual factor is proportional to its probable influence on the outcomes which are of main interest.

    Two dangers to beware of

  6. CFD is sometimes said, by its critics, to be an acronym for Colourful Fluid Dynamics, the allusion being to its employment of realistic-looking images and videos to display its results; and to their power to induce far too much faith in their realism. Such displays should be sparingly used and carefully presented.

  7. The PHOENICS VR front end is good at setting up single-instance runs; but it has no facility for launching parameter-varying multi-runs of the kind which are essential to hazard assessment. Until the deficiency is corrected (as it soon will be by introduction of the PQ1Maker module), hand-editing of the Q1 file is needed; and run-control batch files must be used. The danger is that, the necessary skills being rare, only single-instance runs are made; and far too few of them.

2. Establishing the customer's needs

Properly establishing the customer's needs is not an easy task; for he may have little understanding of the limitations of CFD, or indeed of the independent-of-CFD principle (a) above.

Sometimes the customer may be seeking to provide evidence that his equipment or process satisfies some authority-set quantitative criterion. Error bounds are clearly relevant in such cases.

Alternatively the target may itself be subject to uncertainty, as when the inflammability limits of mixtures of some gas released into air are uncertain. And in such cases it behoves CHAM to make clear that the magnitude of the fluctuations of fuel-air ratio should be calculated; for a spark may ignite an briefly-existing pocket of mixture which differs in composition from the time-average.

In variably the customer has a limited budget to spend; and he will expect to be provided by CHAM with the most informative CFD simulations which can be created within that budget.

The customer must inform CHAM about:

Then CHAM should consider, and advise the customer about: Thereafter an informed discussion can take place, leading to agreement as to what will give the customer best value for money.

It is important that input-range, grid-fineness and reporting options should be costed separately, so that both sides can see that the best value is usually provided by many runs, modest fineness, and minimal reporting.

Examination of recent hazard-assessment projects performed by CHAM reveals that decisions have been arrived at which entailed few runs, excessive fineness, and extravagant reporting.

Evidently, the above described principles have not been observed; and 'best value for money' has not been provided.

3. How the investigation should proceed

As has been mentioned above (item d.) the VR Editor does not lend itself to the making of the connected multi-runs which hazard assessments require. Therefore, although it may be convenient to employ VRE (perhaps with SPPNAM=FLAIR) to create the very first Q1, this file must be parameterised before the hazard-assessment runs can begin.

PQ1Maker will automate at least the simpler stages of this process; but, until it is released for general use, the VRE-written Q1 must be edited by hand. This is not difficult, requiring knowledge of only the rudiments of PIL; and these are easily learned by copying fragments of existing PQ1s.

Working with PQ1s entails using VRE thereafter for only the visual display of the scenarios beng studied. and decidedly not for entering and saving new data. The reason is that VRE does not understand the advanced PIL which parameterisation employs; and indeed is liable to over-write irrecoverably what the human editor has just laboriously composed.

The computations of which results will be presented below have all been conducted by way of a Q1 file originally created by VRE-FLAIR. The runs have taken no more than a few minutes in an HP net-book.

This is how hazard-assessment studies should always be carried out: with many short runs, and much reality-based thinking. Reporting should also be similarly minimal, with just enough visual to support the conclusions which are drawn.

4. Some facts about turbulence

It is well-known to all experienced CFD practitioners that turbulent flows are represented only approximately by its so-called 'turbulence models. These compute time-average values and cannot do justice to their true nature, typefied by the video hyperlinked here.

Emissions into turbulent air near buildings waver throughout building-sized volumes.  This entails that coarse-grid prediction may be more realistic than fine-grid one, rather than less so.

CHAM personnel need to understand this themselves. Moreover, they must convey this knowledge to their customers, and make sure that its practical implications are understood.

This is a professional duty, both to the customer and to CHAM. When it is not performed, CHAM can be held legally responsible for damage resulting from the customer's too-credulous reliance on CHAM's unqualified predictions.

4. Numerics

It is also well-known that the quantitative accuracy, and so the practical reliability, of CFD-based predictions depend greatly on the fineness of the computational grid.

This too needs to be conveyed to customers, not least so as to make clear that sufficiently-fine-for-accuracy grids may be unaffordably expensive.

CHAM's proposal writers need to understand it, too; else CHAM finds itelf committed to the making of runs which cost far more than the customer is being asked to pay.


The message to be conveyed to customers is therefore that, although CFD is good at predicting trends, it can rarely predict absolute values with better than +/- 50% accuracy.

Grid coarsened runs are essential in consultancy

The just-alluded-to facts lead logically to this section heading. Thus: It follows that means must be devised of coarsening the fine grids which are being used to define a particular scenario without losing the essential features of that scenario.

Such a means does exist; and it should be used in all future hazard-assessment consultancies. While not yet available by way of any GUI, it can be activated by simple hand-editing of the so-called OBJINF files printed automatically by the Satellite. It allows the number of intervals in each grid region to be diminished, while the number and position of the regions remains the same. It has been used in the calculations of which results are presented below.

Among the purposes of the exploratory runs is to distinguish the more-important of the scenario-defining attributes from the less-important ones; and one way of doing so is to examine the influences of each attribute separately.

This will now be illustrated by considering a steady-state pollutant- injection situation with no buildings present at all.

The influence of wind direction

The next two image show the predicted positions of the surface on which the mass fraction of HCl is 1.e-5 with wind directions of 90 degrees on the left and 45 degrees on the right.

Comparison of the the images shows that CFD has indeed predicted the trend correctly: the plume of pollutant has changed direction by 45 degrees.

However the maximum value has changed in the ratio 12/8.4; and lengths and widths of the plumes are radically different. For this there is no physical reason; but there is a numerical one. It is called 'numerical diffusion'; and it has been known since the earliest days of CFD [Gosman et al, 1969, p.132] that its magnitude is of the order of the cell dimension times velocity times the sine of twice the angle of inclnation of the velocity vector. Its value is often greatly in excess of the magnitude of the effective diffusivity calculated by the turbulence model.

This conclusion can be tested by performing a run with no turbulence model at all, i.e. for laminar flow. The following image shows the result.

The plume shape is almost the same, and the maximum value has risen only from 8.40 to 8.86 .

One wonders how many results, which CHAM has confidently presented to its customers, have been similarly dominated by numerical diffusion.

Several further remarks are in order relevant to the purpose of the present document, namely:

The influence of choice of turbulence model

That turbulence models disagree with each other is well-known. For long-forgotten reasons the Kim-Chen variant of the k-epsilon model is the default choice of PHOENICS-FLAIR. However, for flows along smooth surfaces such as the ground in the present problem, the LVEL model is likely to fit reality more closely, especially if the grid is coarse. It is therefore interesting to ask how different would be the result if it had been chosen; of course for the 90-degree wind direction so as to minimise numerical diffusion.

The next images show the Kim-Chen and LVEL solutions side-by-side.

The difference in the maximum value is around 10%; but the difference in lateral spread is much greater.

Probably, with a finer grid, the difference would be smaller; but it is clear that the choice of model is likely always to be influential.

The influence of grid fineness

The abiliy to hand-edit the infob2 file allows local grid refinement, which is economically feasible, instead of whole-domain refinement, which is not. In the first (left) image below, the number of intervals in the z-direction region nearest to the ground has been increased five-fold. In the second (image) the same has been done in x- and y-direction regions just downwind of the pollutant injection point.

The maximum pollutant concentration has increased nearly five-fold in both images. This is understandable because the material is being supplied at the ground surface into cells which now receive one fifth of the previous amount of air.

The lesser lateral spread of plume in the right-hand image is a consequence of the reduced numerical diffusion coefficient, this being proportional, as stated above to the size of cell.

The influence of the wind speed

The postulated speedof the wind is of obvious importance in atmospheric-pollution problems; for, the lower it is, the higher will be the concentration of the pollutant. The following two images show the change of shape of the 1.E-4 pollutant-concentration surface downstream of the source, for two wind speeds, respectively 2.1 and 1.9 m/s.

Evidently, the 10% reduction in wind speed has led to a rise of maximum concentration of about 10% and to a corresponding enlargement of the volume of polluted air. ---------------------------------------------- To be continued