10th World Wide Workshop for Young Environmental Scientists 31 May - - PowerPoint PPT Presentation

Identification of the unknown pollution Identification of the unknown pollution source in the Alsatian aquifer (France) source in the Alsatian aquifer (France) through groundwater through groundwater modelling modelling and and Artificial Neural Networks applications Artificial Neural Networks applications

10th World Wide Workshop for Young Environmental Scientists

• Ing. Maria Laura Foddis

31 May - 4 June 2010

University of Cagliari – Italy Department of Land Engineering (DIT) Section of applied Geology and applied geophysics University of Strasbourg – France Laboratory of Hydrology and Geochemistry

• f Strasbourg (LHyGeS)

INTRODUCTION

Pollution may results from contamination whose origins are generated at different times and places where these contaminations have been actually found.

Such situations needs to develop techniques that allow identify unknown contaminant sources behaviour in time and space The purpose of this work aims at studying the spreading of a dangerous chemical - carbon tetrachloride (CCl4 ) - that contaminated, due to a tanker accident in 1970, a part of the largest aquifer in Western Europe:

The Alsatian aquifer (France)

The exact amount of the chemical infiltrated is unknown and this constitutes the main issue for its individuation and remediation.

10th World Wide Workshop for Young Environmental Scientists 31 May - 4 June 2010

The objective of the research is to find a solution of the inverse problem for the Alsatian aquifer: using known contamination concentration data from pumping wells behaviour and temporal evolution of the unknown pollution source is reconstructed

10th World Wide Workshop for Young Environmental Scientists 31 May - 4 June 2010

OBJECTIVE OF THE RESEARCH

GEOGRAPHICAL LOCATION OF THE ALSATIAN AQUIFER

Upper Rhine Graben valley

10th World Wide Workshop for Young Environmental Scientists 31 May - 4 June 2010

Alsatian Region

STRUCTURE OF THE AQUIFER

10th World Wide Workshop for Young Environmental Scientists 31 May - 4 June 2010

Continental extensive alluvial aquifer

Phreatic aquifer fed by meteoric deposits and drained mainly by rivers and human activities. It has a Structure layered with a random superposition of different alluviums

Monitoring wells Pumping wells Monitoring wells Pumping wells

HYSTORY OF THE AQUIFER POLLUTION BY CCl4

11 December 1970 - accident An unknown quantity of carbon tetrachloride (CCl4 ) spreading in the accident area ~ 4m3 (data SGAL -

Service Géologique d’Alsace-Lorraine)

CCl4 62,4 ÷ 56,2 μg/l 1991 – Erstein – first analyses carried out by BRGM

(BRGM : Bureau de Recherche Géologique et Minière)

These quantities exceeded the safe limits recommended by the World Health Organization 2 μg/l This high level of CCl4 concentration has caused serious problems in the region by contaminating the most important drinking water source in the area

10th World Wide Workshop for Young Environmental Scientists 31 May - 4 June 2010

Volatile Organic Chemical does not naturally occur in the environment. It is miscible with most aliphatic solvents but has low solubility in water.

DNAPL – Dense Non Aqueous Phase Liquids

Molecular weight 153,84 g/mol Concentration limit that permit the perception of its sweet smell in water 0,52 mg/l Boiling point 76,6 °C Solubility at 20°C 800 mg/l Density 1,59 mg/l at 20°C Koc – Soil Sorption Coefficient 71 Vapor pressure 91,3 mm Hg a 20°C Henry’s Law Constant 3,04*10-2 atm-m3/mol a 24,8°C

In Alsace aquifer CCl4 and its toxic constituents have actions of:

10th World Wide Workshop for Young Environmental Scientists 31 May - 4 June 2010

PHYSICAL AND CHEMICAL PROPERTIES OF CCl4

Volatilization Sorption

Convection Dispersion Diffusion

Are insignificant:

DIFFICULTIES OF THE PROBLEM SOLUTION

• The exact amount of the chemical infiltrated and the source

behaviour and how the pollutant feeds the contamination is unknown.

• Carbon tetrachloride has low solubility in water, it represents a

continuous source of contamination for the groundwater.

• High uncertainty in the aquifer formation because of its

heterogeneity and its different permeable layers.

10th World Wide Workshop for Young Environmental Scientists 31 May - 4 June 2010

• The chemical properties such as the solubility in water, diffusion,

volatilization, and degradation coefficients are uncertain.

MATERIAL & METHODS

The ANNs are used in processing information. These have the ability to solve complex problems and the capacity to approximate any input-output relationship. This patterns are based on flux and transport model in porous media of CCl4 contamination in the studied domain.

10th World Wide Workshop for Young Environmental Scientists 31 May - 4 June 2010

Training set Validation set Test set

The inverse problem for the Alsatian aquifer is solved using Artificial Neural Network (ANN)

The Patterns for ANN 26 different scenarios of source contamination behaviour are performed with a 3D model of the Alsatian aquifer created by Fluid and Solid Mechanics Institute

• f Strasbourg using TRACES SOFTWARE.

TRACES Transport or Radio Activer Elements in the Subsurface Hoteit et Ackerer, 2003

ALSATIAN AQUIFER NUMERICAL MODEL

10th World Wide Workshop for Young Environmental Scientists 31 May - 4 June 2010

Distribution of CCl4 concentration after 5 years of the accident Tempo di simulazione: 5 anni Tempo di simulazione: 5 anni

Time of simulation: 20000 days ~ 54 years Time of source activity: 11520 days~ 32 years

ALSATIAN AQUIFER NUMERICAL MODEL

10th World Wide Workshop for Young Environmental Scientists 31 May - 4 June 2010

Distribution of CCl4 concentration after 10 years of the accident Tempo di simulazione: 10 anni Tempo di simulazione: 10 anni

ALSATIAN AQUIFER NUMERICAL MODEL

10th World Wide Workshop for Young Environmental Scientists 31 May - 4 June 2010

Distribution of CCl4 concentration after 21 years of the accident Tempo di simulazione: 21 anni Tempo di simulazione: 21 anni

ALSATIAN AQUIFER NUMERICAL MODEL

10th World Wide Workshop for Young Environmental Scientists 31 May - 4 June 2010

Distribution of CCl4 concentration after 28 years of the accident Tempo di simulazione: 28 anni Tempo di simulazione: 28 anni

ALSATIAN AQUIFER NUMERICAL MODEL

10th World Wide Workshop for Young Environmental Scientists 31 May - 4 June 2010

Distribution of CCl4 concentration after 54 years of the accident Tempo di simulazione: 54 anni Tempo di simulazione: 54 anni

The examples obtained with TRACES consisted of matrix of size [m,n]

10th World Wide Workshop for Young Environmental Scientists 31 May - 4 June 2010

PATTERN CONSTRUCTION FOR THE ANN

Matrices were too large to be processed through the ANN

Input matrices are composed of 4 columns: one for each layer in the source. Output matrices are composed of 45 columns: one for each well. In both matrices, rows represent time. For each example we calculated: 2-D discrete Fourier transform (FFT) Among the frequency components of the FFT only the most significant were considered, and the remaining ones were set to zero patterns are normalized in the range [-1 ; +1]

10th World Wide Workshop for Young Environmental Scientists 31 May - 4 June 2010

To construct the ANN several trials were performed to define:

number of neurons 11-11-56 number of hidden layers 1 number of training epochs 100 activation functions for each layer The ANN are developed with the Neural Network Toolbox of MATLAB 7.1 During the training the connection weights were modified in order to minimize the error

• n the training set.

At the same time the error was calculated also on a validation set, independent from the training set. As the validation error began to rise, the training process is interrupted.

TRAINING OF THE NEURAL NETWORK MODEL

⎪ ⎩ ⎪ ⎨ ⎧ = + ⋅ = = + ⋅ u b h W y h y b x W

2 2 1 1

) ( layer Output layer Hidden layer Input σ

Algebraic equations systems

• f the trained ANN

that realize the relationship between input and output patterns

10th World Wide Workshop for Young Environmental Scientists 31 May - 4 June 2010

INVERSION OF THE NEURAL NETWORK MODEL

( )

( )

2 2 1 2 2

b u W W W h

T T

− ⋅ ⋅ ⋅ =

−

( )

h y

1 −

= σ

1 1 1

b W x ⋅ =

−

Once the training phase was completed and all the weights were determined, the inversion of the network could be performed. Knowing the output of the ANN, which derived from a set of measurements in the wells, the corresponding input could be calculated. On the basis of the third equation showed: started by the know output u were determined the vectors h the inverted ANN input pattern x. the vectors y

10th World Wide Workshop for Young Environmental Scientists 31 May - 4 June 2010

INVERSION OF THE NEURAL NETWORK MODEL

The real and the inverted input patterns in frequency domain. There is a strong correspondence between the two patterns.

After the training procedure, the ANN, developed for the definition

• f

the groundwater pollution source that contaminated the Alsatian aquifer, has allowed us to generalize the information contained in the training set. On the basis of the output of the ANN the inverted ANN input pattern x in frequency domain could be determined. The corresponding input in time domain can be calculated by backward applying the pre-processing of the input data.

10th World Wide Workshop for Young Environmental Scientists 31 May - 4 June 2010

CONCLUSIONS

Therefore the development of an ANN able to predict the source behaviour can be useful to design the necessary remediation. These methods are quite inexpensive to develop and could be replicated in developing countries having similar problems of contamination. In developing countries and developed countries, situations of pollution like the case of Alsatian Aquifer are likely to happen.

10th World Wide Workshop for Young Environmental Scientists 31 May - 4 June 2010

BENEFITS AND LIMITATIONS FOR DEVELOPPING COUNTRIES

However this methodologies requires the construction of a coherent number of patterns to adequately describe the input-

• utput relationship.

Thank you for your attention

Identification of the unknown pollution Identification of the unknown pollution source in the Alsatian aquifer (France) source in the Alsatian aquifer (France) through groundwater through groundwater modelling modelling and and Artificial Neural Networks applications Artificial Neural Networks applications

10th World Wide Workshop for Young Environmental Scientists

• Ing. Maria Laura Foddis

31 May - 4 June 2010

University of Cagliari – Italy Department of Land Engineering (DIT) Section of applied Geology and applied geophysics University of Strasbourg – France Laboratory of Hydrology and Geochemistry

• f Strasbourg (LHyGeS)

INTRODUCTION

Pollution may results from contamination whose origins are generated at different times and places where these contaminations have been actually found.

Such situations needs to develop techniques that allow identifying unknown contaminant sources behaviour in time and space The purpose of this work aims at studying the spreading of a dangerous chemical - carbon tetrachloride (CCl4 ) - that contaminated, due to a tanker accident in 1970, a part of the largest aquifer in Western Europe:

The Alsatian aquifer (France)

The exact amount of the chemical infiltrated is unknown and this constitutes the main issue for its individuation and remediation.

10th World Wide Workshop for Young Environmental Scientists 31 May - 4 June 2010

The objective of the research is to find a solution of the inverse problem for the Alsatian aquifer: using known contamination concentration data from pumping wells behaviour and temporal evolution of the unknown pollution source is reconstructed

10th World Wide Workshop for Young Environmental Scientists 31 May - 4 June 2010

OBJECTIVE OF THE RESEARCH

GEOGRAPHICAL LOCATION OF THE ALSATIAN AQUIFER

Upper Rhine Graben

Average extension: width of about 40 km length of about 300km (to Basel from Frankfurt) Thickness of the rift :

• maximum~ 200m
• average in the region ~ 80m

10th World Wide Workshop for Young Environmental Scientists 31 May - 4 June 2010

Alsatian Region and Alsatian Aquifer Map

STRUCTURE AND DIMENSION OF THE ALSATIAN AQUIFER

Surface: 3000km2

Groundwater reserves: about 50 billion m3 Annual renewal: about 1,3 billion m3 Annual exploitation: about 0,5 billion m3

These number are extracted from the note of CIENPPA (Banque des Donnees Eau Modele Hydrodynamique Regional)

Altitude: from +250m in Basel to +130m in Lauterbourg Volume of alluvium:

about 250 billion m3

Tertiary and Quaternary sediments 10th World Wide Workshop for Young Environmental Scientists 31 May - 4 June 2010

Continental extensive alluvial aquifer

Phreatic aquifer fed by meteoric deposits and drained mainly by rivers and human activities. It has a Structure layered with a random superposition of different alluviums (clay, sand, gravels, coarse…)

HYDROGRAPHICAL SISTEM OF THE ALSACE PLAIN

Monitoring wells Pumping wells

Main rivers : Channeled Rhine

Q = 700 ÷ 1500 m3s-1

Ill

Q = 5 ÷ 10 m3s-1

10th World Wide Workshop for Young Environmental Scientists 31 May - 4 June 2010

HYSTORY OF THE AQUIFER POLLUTION BY CCl4

11 December 1970 - accident An unknown quantity of carbon tetrachloride (CCl4 ) spreading in the accident area ~ 4m3 (data SGAL -

Service Géologique d’Alsace-Lorraine)

CCl4 62,4 ÷ 56,2 μg/l 1991 – Erstein – first analyses carried out by BRGM

(BRGM : Bureau de Recherche Géologique et Minière)

These quantities exceeded the safe limits recommended by the World Health Organization 2 μg/l This high level of CCl4 concentration has caused serious problems in the region by contaminating the most important drinking water source in the area

10th World Wide Workshop for Young Environmental Scientists 31 May - 4 June 2010

Volatile Organic Chemical does not naturally occur in the environment. It is miscible with most aliphatic solvents but has low solubility in water.

DNAPL – Dense Non Aqueous Phase Liquids

Molecular weight 153,84 g/mol Concentration limit that permit the perception of its sweet smell in water 0,52 mg/l Boiling point 76,6 °C Solubility at 20°C 800 mg/l Density 1,59 mg/l at 20°C Koc – Soil Sorption Coefficient 71 Vapor pressure 91,3 mm Hg a 20°C Henry’s Law Constant 3,04*10-2 atm-m3/mol a 24,8°C

In Alsace aquifer CCl4 and its toxic constituents have actions of:

10th World Wide Workshop for Young Environmental Scientists 31 May - 4 June 2010

PHYSICAL AND CHEMICAL PROPERTIES OF CCl4

Volatilization Sorption

Convection Dispersion Diffusion

Are insignificant:

DIFFICULTIES OF THE PROBLEM SOLUTION

• The exact amount of the chemical infiltrated and the source

behaviour and how the pollutant feeds the contamination is unknown.

• Carbon tetrachloride has low solubility in water, it represents a

continuous source of contamination for the groundwater.

• High uncertainty in the aquifer formation because of its

heterogeneity and its different permeable layers.

10th World Wide Workshop for Young Environmental Scientists 31 May - 4 June 2010

• The chemical properties such as the solubility in water, diffusion,

volatilization, and degradation coefficients are uncertain.

MATERIAL & METHODS

The ANNs are used in processing information. These have the ability to solve complex problems and the capacity to approximate any input-output relationship. This patterns are based on flux and transport model in porous media of CCl4 contamination in the studied domain.

10th World Wide Workshop for Young Environmental Scientists 31 May - 4 June 2010

Training set, Validation set, Test set.

Artificial Neural Network (ANN)

The inverse problem for the Alsatian aquifer is solved using ANN The Patterns for ANN

TRACES

Transport or RadioActiver Elements in the Subsurface FORTRAN 95

Hoteit et Ackerer, 2003

Fluid and Solid Mechanics Institute of Strasbourg.

combines the mixed-hybrid finite element (MHFE) and discontinuous Galerkin (DG) methods to solve the hydrodynamic state and mass transfer problems

10th World Wide Workshop for Young Environmental Scientists 31 May - 4 June 2010

PATTERN CONSTRUCTION FOR TRAINING THE ANN

To generate the patterns, different scenarios of source contamination behaviour are performed with a 3D model of the Alsatian aquifer created by Fluid and Solid Mechanics Institute of Strasbourg. The model created was calibrated using measured data of carbon tetrachloride concentration collected from 1992 to 2004.

51 Patterns are constructed using TRACES SOFTWARE

N

3D Mesh The domain is discretized into a non uniform mesh with 25388 nodes and 45460 irregular prismatic elements.

NUMERICAL MODEL CHARACTERISTICS

10th World Wide Workshop for Young Environmental Scientists 31 May - 4 June 2010

Geometry of the 3D model Hydraulic conductivity:

variable the vertical between 10-3 and 10-2m/s

Porosity:

variable > 15%.

Hydraulic gradient:

0.7% -0.9%.

Propriety of the aquifer Dimensions: area=6*20Km2

depth =110m

Numbers of layers:

11 layer have different depths

ALSATIAN AQUIFER NUMERICAL MODEL

Time of simulation: 20000 days ~ 54 years Time of source activity: 11520 days~ 32 years

The source is placed into the first 4 layers of the numerical model. The thicknesses

• f the layers are 16, 4, 5, and 5 m from top to bottom.

10th World Wide Workshop for Young Environmental Scientists 31 May - 4 June 2010

Head distribution

Location of the contaminant source Volume of the contaminated aquifer : between 230 m3 and 1300 m3. Surface of infiltration: between 7 and 37 m2.

ALSATIAN AQUIFER NUMERICAL MODEL

10th World Wide Workshop for Young Environmental Scientists 31 May - 4 June 2010

Distribution of CCl4 concentration after 5 years of the accident Tempo di simulazione: 5 anni Tempo di simulazione: 5 anni

ALSATIAN AQUIFER NUMERICAL MODEL

10th World Wide Workshop for Young Environmental Scientists 31 May - 4 June 2010

Distribution of CCl4 concentration after 10 years of the accident Tempo di simulazione: 10 anni Tempo di simulazione: 10 anni

ALSATIAN AQUIFER NUMERICAL MODEL

10th World Wide Workshop for Young Environmental Scientists 31 May - 4 June 2010

Distribution of CCl4 concentration after 21 years of the accident Tempo di simulazione: 21 anni Tempo di simulazione: 21 anni

ALSATIAN AQUIFER NUMERICAL MODEL

10th World Wide Workshop for Young Environmental Scientists 31 May - 4 June 2010

Distribution of CCl4 concentration after 28 years of the accident Tempo di simulazione: 28 anni Tempo di simulazione: 28 anni

ALSATIAN AQUIFER NUMERICAL MODEL

10th World Wide Workshop for Young Environmental Scientists 31 May - 4 June 2010

Distribution of CCl4 concentration after 54 years of the accident Tempo di simulazione: 54 anni Tempo di simulazione: 54 anni

The examples obtained with TRACES consisted of matrix of size [m,n]

10th World Wide Workshop for Young Environmental Scientists 31 May - 4 June 2010

PATTERN CONSTRUCTION FOR THE ANN

Matrices were too large to be processed through the ANN

Input matrices are composed of 4 columns: one column for each layer in the source. Output matrices are composed of 45 columns: one column for each well. In both matrices, rows represent time.

For each example we calculated: 2-D discrete Fourier transform (FFT) Among the frequency components of the FFT only the most significant were considered, and the remaining ones were set to zero patterns are normalized in the range [-1 ; +1]

10th World Wide Workshop for Young Environmental Scientists 31 May - 4 June 2010

To construct an ANN it need to define: number of neurons, number of hidden layers , activation functions for each layer. To do this several trials were performed and we chose: 11 input neurons, 56 output neurons 1 hidden layer with 11 neurons. The ANN are developed with the Neural Network Toolbox of MATLAB 7.1 During the training the connection weights were modified in order to minimize the error on the training set, but at the same time the error was calculated also on a validation set, independent from the training set, and as the validation error began to rise, the training process is interrupted.

TRAINING OF THE NEURAL NETWORK MODEL

10th World Wide Workshop for Young Environmental Scientists 31 May - 4 June 2010

On the basis of preliminary proof: the number of training epochs was set at 100, an 11-11-56 Multi Layer Perceptron (MLP) ANN was trained with a set of 26 examples, the mean squared error of output lower than 10-3.

INVERSION OF THE NEURAL NETWORK MODEL

Algebraic equations systems of the trained ANN that realize the relationship between input and output patterns

⎪ ⎩ ⎪ ⎨ ⎧ = + ⋅ = = + ⋅ u b h W y h y b x W

2 2 1 1

) ( layer Output layer Hidden layer Input σ

10th World Wide Workshop for Young Environmental Scientists 31 May - 4 June 2010

INVERSION OF THE NEURAL NETWORK MODEL

( )

( )

2 2 1 2 2

b u W W W h

T T

− ⋅ ⋅ ⋅ =

−

( )

h y

1 −

= σ

1 1 1

b W x ⋅ =

−

Once the training phase was completed and all the weights were determined, the inversion of the network could be performed. Knowing the output of the ANN, which derived from a set of measurements in the wells, the corresponding input could be calculated. On the basis of the third equation showed: started by the know output u were determined the vectors h the inverted ANN input pattern x. the vectors y

10th World Wide Workshop for Young Environmental Scientists 31 May - 4 June 2010

INVERSION OF THE NEURAL NETWORK MODEL

The real and the inverted input patterns in frequency domain. There is a strong correspondence between the two patterns.

10th World Wide Workshop for Young Environmental Scientists 31 May - 4 June 2010

INVERSION OF THE NEURAL NETWORK MODEL

Real and the inverted input patterns in frequency domain

This precision can be improved by considering that the input patterns are normalized in the range [-1 ; +1], so that by cutting the out-of-range values, the precision is still better.

After the training procedure, the ANN, developed for the definition

• f

the groundwater pollution source that contaminated the Alsatian aquifer, has allowed us to generalize the information contained in the training set. On the basis of the output of the ANN the inverted ANN input pattern x in frequency domain could be determined. The corresponding input in time domain can be calculated by backward applying the pre-processing of the input data.

10th World Wide Workshop for Young Environmental Scientists 31 May - 4 June 2010

CONCLUSIONS

Therefore the development of an ANN able to predict the source behaviour can be useful to design the necessary remediation. These methods are quite inexpensive to develop and could be replicated in developing countries having similar problems of contamination. In developing countries and developed countries, situations of pollution like the case of Alsatian Aquifer are likely to happen.

10th World Wide Workshop for Young Environmental Scientists 31 May - 4 June 2010

BENEFITS AND LIMITATIONS FOR DEVELOPPING COUNTRIES

However this methodologies requires the construction of a coherent number of patterns to adequately describe the input-

• utput relationship.

Thank you for your attention