GNGTS 2021 - Atti del 39° Convegno Nazionale

387 GNGTS 2021 S essione 3.1 REAL-TIME NEURAL NETWORK INVERSION OF AIRBORNE TDEM DATA P. Bai 1 , G. Vignoli 2 , A. Viezzoli 3 , G. Vacca 1 1 Univ. of Cagliari, Italy 2 Univ. of Cagliari, Italy & GEUS, Denmark 3 Aarhus Geophysics ApS, Denmark The possibility of getting reliable results very quickly after, or even during, the data collection would be crucial, not merely for quality check, but also for adjusting the location of the proposed flight lines during an airborne time-domain (ATEM) acquisition. This kind of readiness could have a large impact in terms of maximization of the Value of Information of the measurements to be acquired. Besides, the relevance of fast tools for reconstructing resistivity models from ATEM data is demonstrated by the routine use of Conductivity-Depth Imaging (CDI) methodologies in mineral exploration. In fact, CDIs are extremely efficient from a computational perspective, and, at the same time, they preserve a very high lateral resolution. Hence, they are often preferred to inversion strategies even if the latter approaches are, in general, more accurate in terms of reconstruction of the depth of the targets and of retrieval of true resistivity values. Here, we discuss a novel approach, based on Neural Network (NN) techniques, capable of reconstructing resistivity models with a quality comparable with the inversion strategy, but in a fraction of the time: seconds on a laptop versus hours on a computational server. We demonstrate the advantages of the proposed novel approach on synthetic and field datasets (Bai et al., 2020). Methodologies ATEM data are usually inverted by minimizing an objective functional consisting of a data misfit term plus a stabilizer. Hence, the objective functional to be minimized is often written: twork Inversion of Airborne TDEM data. i, Italy & GEUS, Denmark sics ApS, Denmark Italy liable results very quickly after, or even during, the data collection would be crucial, , but also for adjusting the location of the proposed flight lines during an airborne sition. This kind of readiness could have a large impact in terms of maximization of of the measurements to be acquired. Besides, the relev nce of fast ools for odels from ATEM data i demonstrated by the routine use of Conductivity-Dep h es in min ral exploration. In fact, CDIs are extremely efficient from a computational time, they preserve a very high lateral resolution. Hence, they are often preferred to the latter approaches are, in general, more accurate in terms of reconstruction of the trieval of true resistivity values. Here, we discuss a novel approach, based on Neural capable of reconstructing resistivity models with a quality comparable with the raction of the time: seconds on a laptop versus hours on a computational server. We of the proposed novel approach on synthetic and field datasets (Bai et al., 2020). a are usually inverted by minimizing an objective functional consisting of a data Hence, the objective functional to be minimized is often written: P ( λ ) ( d , m ) = ‖ W d ( d obs − F ( m ) ) ‖ L 2 2 + λ s ( m ) , (1) r of the measurements; (b) m is the vector of the model parameters; (c) F is the connecting the model m to the corresponding data; F takes into account the physics acteristics of the acquisition system; (d) W d is the data weighting matrix usually uncertainty; (e) s ( m ) is the stabilizer incorporating the prior knowledge about the overed; (f) the multiplier λ controls the balance between the data and the prior inistic scheme we are using here to assess the performances of the alternative consider a one-dimensional model parameterization. Hence, m and F rely on the ch individual data sounding, and each associated model, is handled independently re precisely, while the forward modelling F is always one-dimensional, a lateral hboring models is imposed by means of the regularization term. In this respect, oice, we adopt the very common option of s ( m ) equal to the minimum gradient s ( m )= ‖ ∇ m ‖ L 2 2 . (2) uctivity distribution is locally 1D, the stabilizer acts both along the vertical ( z ) and the orcing some level of laterally coherency (without being truly 2D/3D). This is the (1) where (a) Inversion of Airborne TDEM data. GEUS, Denmark , Denmark ults very quickly after, or even during, th data collec ion would be crucial, o for adjusting the location of the proposed flight lines during an airborn his kind of readiness could have a large impact in terms of maximization of measurements to be acquired. Besides, the relevance of fast tools for m ATEM data is demonstrated by the routine use of Conductivity-Depth eral exploration. In fact, CDIs are extremely efficient from a computational ey preserve a very high lateral resolution. Hence, they are often preferred to approac s are, in general, more accurate in terms of reconstruction of the f true resistivity values. Here, we discuss a novel approach, based on Neural of reconstructing resistivity models with a quality comparable with the f the time: seconds on a laptop versus hours on a computational server. We oposed novel approach on synthetic and field datasets (Bai et al., 2020). ually inverted by minimizing an objective functional consisting of a data he objective functional to be minimized is often written: P ( λ ) ( d , m ) = ‖ W d ( d obs − F ( m ) ) ‖ L 2 2 + λ s ( m ) , (1) measurements; (b) m is the vector of the model parameters; (c) F is the ng the model m to the corresponding data; F takes into account the physics cs of the acquisition system; (d) W d is the data weighting matrix usually ty; (e) s ( m ) is the stabilizer incorporating the prior knowledge about the (f) the multiplier λ controls the balance between the data and the prior scheme we are using here to assess the performances of the alternative r a one-dimensional model parameterization. Hence, m and F rely on the idual data sounding, and each associated model, is handled independently sely, while the forward modelling F is always one-dimensional, a lateral models is imposed by means of the regularization term. In this respect, adopt the very common option of s ( m ) equal to the minimum gradient s ( m )= ‖ ∇ m ‖ L 2 2 . (2) distribution is locally 1D, the stabilizer acts both along the vertical ( z ) and the ome level of laterally coherency (without being truly 2D/3D). This is the is the vector of the measurements; (b) l-Time Neural Network Inversion of Airborne TDEM data. ai, Univ. of Cagliari ignoli, Univ. of Cagliari, Italy & GEUS, Denmark iezzoli, Aarhus Geophysics ApS, Denmark acca, Univ. of Cagliari, Italy po sibility of getting eliable res lts very quickly after, or ven during, the data coll ction would be crucial, merely for quali y check, but also for adjusting the location of the p opo d flight lines during an airborne -domain (ATEM) acquisition. This kind of readiness could have a large impact in terms of maximization of Value of Information of the measurements to be acquired. Besides, the relevance of fast tools for nstructing resistivity models from ATEM data is demonstrated by the routine use of Conductivity-Depth ging (CDI) methodologies in min ral exploration. In fact, CDIs are extreme y efficient from a computational pective, and, at the same time, they preserve a v ry high lateral resolution. Hence, they are often preferred to rsion strat gies even if the latter approaches are, in gen ral, more accurate in terms of reconstruction of the h of the targets and of retri val of true re istivity values. Here, we discuss a novel approach, based on Neural ork (NN) techniques, capable of reconstructing resistivity models with a quality comparable with the rsion strategy, but in a fraction of the time: econds on a laptop versus hours on a computational server. We onstrate the advantages of the proposed novel approach on synthetic and field datasets (Bai et al., 2020). hodologi s. ATEM data are usually inverted by minimizing an bjective functional consisting of a data it term plus a stab lizer. Hence, the objective functional to be minimized is often written: P ( λ ) ( d , ) = ‖ W d ( d obs − F ( m ) ) ‖ L 2 2 + λ s ( m ) , (1) re (a) d obs is the vector of the measurements; (b) m is the ve tor of the model parameters; (c) F is the ard modelling operator connecting the model m to the corresponding data; F takes into account the physics e process and the characteristics of the acquisition system; (d) W d is the data weighting matrix usually d on the measurement uncertainty; (e) s ( m ) is the stabilizer incorporating the prior knowledge about the tivity model to be recovered; (f) the multiplier λ controls the balance between the data and the prior rmation. In the deterministic scheme we are using here to assess the performances of the alternative o ch based o NN, we consider a one-d mensio al model parameterization. Hence, m and F rely on the l 1D assumption. So, each individual data s unding, ach associated model, is handled independently the adjacent ones. More pr c sely, while the f rward modelling F is ways one-dimensional, a lateral traint between th neighboring models is imposed by means of the r gularization term. In this respect, erning h stabilizer choice, we adopt the very c mmon option of s ( m ) equal to the minimum gradient stabilizer s ( m )= ‖ ∇ m ‖ L 2 2 . (2) ce, even though the conductivi y distribution is locally 1D, the stabilizer acts both along the vertical ( z ) and the z ntal ( x ) direction, enf rcing some level of lateral y coherency (without being truly 2D/3D). This is the is the vector of the model parameters; (c) n of Airborne TDEM data. mark ickly after, or even during, the data collection would be crucial, ing the loc tion of the proposed flight li s dur g n airborne eadin ss could have a large impact in t rms of maximization of ts to be acquired. Besides, the relevance of fast tool for ata is demons rated by the routine use of Conductivity-Depth tion. In fact, CDIs are extremely efficient from a computational a very high lateral resolution. Hence, they are often pref rred to are, in general, more accurate in ter s of reconstruction of the ity value . Here, we discuss novel approach, based on Neural cti g resistivity m dels with a quality comparable with the ec nds on a laptop ver us hours on a computational erver. We l approach on synthetic and field datasets (Bai et al., 2020). d by minimizing an objective functional consisting of a data functional to be minimized is often written: W d ( d obs − F ( m ) ) ‖ L 2 2 + λ s ( m ) , (1) ts; (b) m is the vect r of the model parameters; (c) F is the l m t the corresponding data; F takes int account the physics uisition system; (d) W d is the data weighting matrix usually ) is the tabilizer incorporating the prior knowledge about the iplier λ co trols the balance between the data and the prior are using here to assess the performances of the alternative nsional model parameteriz tion. Hence, m and F rely on the ounding, and each associated model, is handled independently the forward odelli g F is always one-dimensional, a lateral mposed by means of the regularization term. In this respect, ery common option of s ( m ) equal to the minimum gradient s ( m )= ‖ ∇ m ‖ L 2 2 . (2) s locally 1D, the stabilizer acts both along the vertical ( z ) and the f laterally coherency (without being truly 2D/3D). This is the is the forward modelling operator connecting the model Real-Tim Neural Ne work Inversion of Airborne TDEM data. P. Bai, Univ. of Cagliari G. Vignoli, Univ. of Cagliari Italy & GEUS, Denmark A. V ezzoli, Aarhus Geophysics ApS, Denmark G. Vacca, Univ. of Cagliari, Italy The possibility f getti g reliable results very quickly after, or even during, the da a collection would be not mer ly for quality ch ck, but also for adjusting the location of the propos d flight lines during an time-domain (ATEM) acquisi ion. Thi kind of readiness could have a large impact in terms of maximi the V lue of Information of the measurements to be acquired. Besides, the relevance of fast t r construc ing resistivity models from ATEM data is demonstrated by the routine use of Conductivit Imaging (CDI) methodologi s in mineral exploration. In fact, CDIs are extremely efficient from a comp perspect ve, a d, at the same time, they preserve a v ry high lateral resolution. Hence, they are ften pr inversion strategie ev n if the latter pproaches are, in general, more accurate in terms of reconstructi depth of the t rgets an of r tri val of true resistivity values. Here, we discuss a novel a p oach, based o Network (NN) techniques, capable of reconstructing resistivity models with a quality comparable inversion strategy, but in a fra tion of the time: econds on a laptop v rsus hours on computational se demonstrate t e adv tag s of the proposed novel approach on synthetic and field datasets (Bai et al., 20 Methodologies. ATEM data are usually i ver ed by minimiz ng an objective functional consisting misfit term plus a stabil zer. Hence, the objective functional to be minimized is often written: P ( λ ) ( d , m ) = ‖ W d ( d obs − F ( m ) ) ‖ L 2 2 + λ s ( m ) , where (a) d obs is the vector of the measurements; (b) m is the vector of the model parameters; (c) forward modelling operator conn cting the model m to the corresponding data; F takes into account th of the process and the c aracteris ics of the acquisition system; (d) W d is the data eighting matri based on the me surement uncertainty; (e) s ( m ) is the stabilizer incorporating the prior knowledge a resistivity model to b recovered; (f) the multiplier λ controls the balance between t data and t information. In the deterministic scheme we are using here to assess the performances of the al approach based on NN, we consider a one-dimensional model parameterization. Hence, m and F rel local 1D assumpti n. So, each individual ata soun ing, and each associated model, is handled indep from the adjace t ones. M re precisely, while the forward odelling F is always one-dimensional, constraint between the neighbor ng models is imposed by means of the regularization term. In this concerning the stabilizer choice, we adopt the very common option of s ( m ) equal to the minimum norm stabilizer s ( m )= ‖ ∇ m ‖ L 2 2 . Hence, even though the conductivity distribu ion is locally 1D, the stabilizer acts both along the vertical ( z horizontal ( x ) direction, enforcing some level of laterally coherency (without being truly 2D/3D). Th to the corresponding data; Real-Time Neural Network Inversion of Airborne TD P. Bai, Univ. of Cagliari G. Vignoli, Univ. of Cagliari, Italy & GEUS, Denmark A. Viezzoli, Aarhus Geophysics ApS, Denmark G. Vacca, Univ. of Cagliari, Italy The po ibili y of g tting reliable results very quickly aft r, or even uri not merely for quality c e k, but also for adjusting t e cation of the tim -domain (ATEM) acquisition. This kind of readiness could have a l th Value of Information of th measurements to be acquired. Be reconst ucting resistivity models f o ATEM data is demonstrated by Imaging (CDI) methodologies in mineral exploration. In fact, CDIs are e erspective, and, t the same time, they preserve a very high lateral resol inv ion s ra egies ev n if the latter approach s are, in general, more a de th of the t rgets and of retrieval of true resistivity v lues. Here, we di Network (NN) t chniques, capable f reconstructing resistivity mod inversion s rat gy, but in fractio o the time: seconds n a lapt p vers d monstrate the advantages of the proposed novel approach on synthetic Methodologies. ATEM data are usually i ve ted by minimizing an o misfit term plus a stabilizer. Hence, the obj ctive fu ctional to be minimi P ( λ ) ( d , m ) = ‖ W d ( obs − F ( m ) ) ‖ L 2 2 + λ w ere (a) d obs is the vector of the meas rem nts; (b) m is th vector forward modelling operator connecti g the model m to the correspondin of he process and the characteristics of the acquisition sys em; (d) based o the measure e t uncert inty; (e) s ( m ) is the stabilizer inc r sistivity model to be rec vered; (f) the multiplier λ controls the b information. In the deterministic sch me we are using here to asse approach based o NN, w consider a one-dimensional model param t local 1D assumption. So, each individual data sounding, and each asso fr m the adjac nt ones. More precisely, while the forwar odelling co straint between the neighboring models is imposed by means of t conc rning the stabilizer ch ice, we adopt the very common option o norm stabilizer s ( m )= ‖ ∇ m ‖ L 2 2 . He ce, ven though the conductivity distribution is locally 1D, th stabili ho iz ntal ( x ) direction, enforcing some level of laterally coherency ( takes into account the physics of t e process and th chara te istics of the acquisition system; (d) nversion of Airborne TDEM data. EUS, Denmark Denmark lts very quickly after, or even during, he data collection would be crucial, for adjusting the location of the proposed flight lines during an airborne s kind of readin ss could have a arge impact in terms of maximization of easurements to be acquired. Besides, the relevance of fast to l for ATEM data is demonstrated by the routine use of Conductivity-Depth ral exploration. I fact, CDIs are ex remely efficient from a computational y preserve a very high lateral resolution. Hence, they are often preferred to pproaches are, i general, more accurate in terms of reconstruction of the true resistivity values. Here, we discuss a novel approach, based on Neural f recon tructing resistivity models with a quality comparable with the th time: seconds on a laptop versus hours on a computational server. We osed novel approach on synthetic and field datasets (Bai et al., 2020). lly inverted by minimizing an objective functional consisting of a data objective func ional to be inimized is often written: λ ) ( d , m ) = ‖ W d ( d obs − F ( m ) ) ‖ L 2 2 + λ s ( m ) , (1) easurements; (b) m is the vector of the model parameters; (c) F is the the model m to the corresponding data; F takes into account the physics of the acquisition system; (d) W d is the data weighting matrix usually ; (e) s ( m ) is the stabilizer incorporating the prior knowledge about the ) the multiplier λ controls the balance between the data and the prior eme we are using here to assess the performances of the alternative a one-dimensional model parameterization. Hence, m and F rely on the ual data sounding, and each associated model, is handled independently ly, while the forward modelling F is always one-dimensional, a lateral odels is imposed by means of the regularization term. In this respect, adopt the very common option of s ( m ) equal to the minimum gradient s ( m )= ‖ ∇ m ‖ L 2 2 . (2) stribution is locally 1D, the stabilizer acts both along the vertical ( z ) and the is the data w ighting matrix usually based on the measur ment uncertainty; (e) Real-Time Neur l Network Inversion of Airborne TDEM data. P. Bai, Univ. of Cagliari G. Vignoli, Univ. of Cagliari, Italy & GEUS, Denmark A. Viezzoli, Aarhus Geophysics ApS, Denmark G. Vacca, Univ. of Cagliari, Italy The possibility of get ing reliable r sults very quickly af er, or even during, the data coll not merely for quality check, but also for adjusting the location of the p oposed flight l time-domain (ATEM) acquisition. This kind of readiness could have a large impact in te the Value of Information of the measurement t be acquired. Besides, the relev reconstructing resistivity models from ATEM data is demonstrated by the routine use Imaging (CDI) methodolog es in miner l explora o . In fact, CDIs are extremely efficie perspective, nd, at the same time, they preserv a very high lateral resolution. Hence, th inversion strategies eve if the latter approaches are, in general, more accurate in terms depth of the target and f retrieval of true resistivity values. Here, we discuss a novel ap Network (NN) techniques, capable of reconstruc resistivity models with a qualit i version strategy, b t in a fraction of the time: seconds on a laptop versus hours on a co demonstrat the a vantages of he proposed novel approach on synthetic and field dataset Methodologies. ATEM data are usually inverted by minimizing an objective function misfit term plus a stabilizer. Hen e, the objective functional to be m imized is often writ P ( λ ) ( d , m ) = ‖ W d ( d obs − F ( m ) ) ‖ L 2 2 + λ s ( m ) , where (a) d obs is the vector of the measurements; (b) m is the vector of the model p forward mo elling operator con ecting the model m to the corresponding data; F takes i of th process and t charact ristics of the acq isition system; (d) W d is the data w based on the measurement uncertainty; (e) s ( m ) is the stabilizer incorporating the pri resistivity model to be recovered; (f) the multiplier λ controls the balance between informa ion. In deterministic scheme w are using here to assess the performa approach based on NN, we consider a one-dime sional model parameterization. Hence local 1D assumption. S , ach individual data sou ding, and each associated model, is f om the adjacent ones. More precisely, while the forward modelling F is always on constrai t between the neighboring models is imposed by means of the regularizatio concerning the stabiliz r choice, we adopt the very common option of s ( m ) equal to norm stabilizer s ( m )= ‖ ∇ m ‖ L 2 2 . Hence, even though the conductivity dis ribution is locally 1D, the stabilizer acts both alon is the stabilizer incorporating the prior knowledge about the resistivity model to be recovered; (f) he multipli r ion of Airborne TDEM data. enmark k quickly after, or even during, the data c llection would be crucial, sting the location of the proposed flight lines during an airborne f readiness could have a large impact in terms of maximization of ents to be acquired. Besides, the relevance of fast tools for data is demonstrated by the routine use of Conductivity-Depth ration. In fact, CDIs are extremely efficient from a computational e a very high lateral resolution. Hence, they are often preferred to es are, in general, more accurate in terms of reconstruction of the stivity values. Here, we discuss a novel approach, based o Neural structing resistivity models with a quality comparable with the : seconds on a laptop versus hours on a computational server. We vel approach on synthetic and field datasets (Bai et al., 2020). rted by mi imizing an objective functional consisting of a data ve functi nal to be minimiz d is often written: = ‖ W d ( d obs − F ( m ) ) ‖ L 2 2 + λ s ( m ) , (1) ents; (b) m is the ve t r of the model param ters; (c) F is the del m to the corresponding data; F takes into accou t t physic acquisition system; (d) W d is the data weighting matrix usually m ) is the stabilizer incorporating the prior knowledge about the ultiplier λ controls the balance between the data and the prior e are using here to assess the performances of the alternative mensional model parameterization. Hence, m and F rely on the sounding, and each associated model, is handled independently e the forward modelling F is always one-dimensional, a lateral s i posed by means of the regularization term. In this respect, e very common option of s ( m ) equal to the minimum gradient s ( m )= ‖ ∇ m ‖ L 2 2 . (2) controls the balance betwe n the data a d the prior information. In the deterministic scheme we are using here to assess the performances of the alternative approach bas d n NN, we consid r a n -dimens onal model p r m terization. Hence, Real-Time Neural Network Inversion of Airborne TDEM data. P. Bai, Univ. of Cagliari G. Vignoli, Univ. of Cagliari, Italy & GEUS, Denmark A. Viezzoli, Aa hus Geophysics ApS, Denmark G. Vac a, Univ. of Cagliari, Italy The possibility of getting liable results very quickly fter, or even during, the data co not merely for q ality ch ck, but als for a justing the location of the proposed fligh time-domain (ATEM) acquisition. This kind of readiness could have a large impact in the Value of Information of the measurements to be acquired. Besides, the rel reconstructing resistivity models from ATEM data is demonstrated by the routine u Imaging (CDI) methodologies in mineral exploration. In fact, CDIs are extremely effici perspective, and, at the same time, they preserve a very high lateral resolution. Hence, t inversion strat gies even if the latter approach s are, in general, more accurate in term depth of the targets and of retrieval of true resistivity values. Here, we discuss a novel a Network (NN) techniques, capable of reconstructing resistivity models with a qual inversion strategy, but in a fraction of the time: seconds on a laptop versus hours on a demonstrate the advantages of the proposed novel approach on synthetic and field datas Methodologies. ATEM data are usually inv rted by minimizing an objective functi misfit term plus a stabiliz r. Hence, h objective functional be mi imized is often w P ( λ ) ( d , m ) = ‖ W d ( d obs − F ( m ) ) ‖ L 2 2 + λ s ( m ) , where (a) d obs is the vector of th measure ents; (b) m is the vector of the model forward modelling operator connecti g t e model m to the corresp ding data; F take of the process and the characteristics of the acquisition system; (d) W d is the d ta ased o the measurement uncertainty; (e) ( m ) is he stab lizer incorporating th p resistivity model to be recovered; (f) the multipli r λ con rols the balance betwee information. In the eterministic sch me we are using here to assess the perfor approach based on NN, we consider a one-dimensional model parameterization. Hen local 1D assumption. So, each individual data sounding, and each associated model, rom the adjacent ones. More precisely, while the forward modelling F is always constraint between t neighboring m dels is imposed by means of the regularizati concerning the stabilizer choice, w dopt t ver common option of s ( m ) equal norm stabilizer and Real-Tim eural Network Inversion f Airborne TDE P. Bai, Un v. of Cagliari G. Vignoli, Univ. of Cagliari, Italy & GEUS, Denmark A. Viezzoli, Aarhus Ge hysics ApS, Denmark G. Vacca, Un v. of Cagliari, Italy The p ssibility of getting re iab e r sults very qui kly after, or even during, t not mere y f r quality check, but also for adjusting the location of the prop t me-domain (ATEM) acqu ition. This kind of readiness could have a large the Value of I f rmation of the measur ments to be acquir d. Beside reconstructing resistivity odels from ATEM data is d mons rated by the Imagi g (CDI) methodologies in mineral explor tion. In f ct, CDIs are xtre perspective, and, at th same time, th y pr serve a very high lateral resolutio i version strategies even if the latter approaches are, in general, more accur depth of the targets and of retrieval of true resistivity values. Here, we discus etwork (NN) te hniqu s, capable of reconstructing resistivity models w nversion strategy, but i a fraction of the time: econds on a laptop versus h demonstrate the dvantages of the proposed novel approach on synth tic nd Methodologies. ATEM data are usually invert d by minimizing an object misfit erm plus a s abilizer. Hence, the o jective functional to b minimized P ( λ ) ( d , m ) = ‖ W d ( d obs − F ( m ) ) ‖ L 2 2 + λ s ( where (a) d obs is the vector of the measurement ; (b) m is the vector of t forward modelling p ra r connecting the model to th correspo ing da of the process and the characteristics of the acquisition system; (d) W is based on the me sur ment uncertainty; (e) s ( m ) s th stabilizer cor ora resistivi y model to be recovered; (f) th multiplier λ controls the bala information. In the deterministic scheme we ar using here o assess t approach based on NN, we c si er a one-dimensional model p rameteriz local 1D assumption. So, each individu l dat sounding, nd e ch ass ciat from the adjacent ones. More precisely, while the f rward modelling F i c nstrai t between the neighboring models is impo ed by means of the r c n erning the stabilizer choice, w adopt th ery common option of s norm stabilizer rely on the local 1D ssumpti n. So, each individual data sounding, and ac ssociat d m del, is handled independently fro the adjacent nes. M re precisely, while the forward modelling Real-Time Neura Network Inversion of Airborne TD P. Bai, Univ. of Cagliari G. Vignoli, Univ. of Cagliari, Italy & GEUS, Denmark A. Viezzoli, A rhus Ge physics ApS, Denmark G. Vacc , Univ. of Cagliari, Italy The possibility of gett ng r liab e r sult very quickly after, or even duri not merely for quality check, but also for adjus ing the location of the time-domain (ATEM) acqu ition. This kind of readiness could have a l the Va ue f Information of the m asurements to be acquired. B r construc ing resistivity models from ATEM data is demonstrated by Imaging (CDI) methodologies in iner l exploration. I fact, CDIs are e perspective, and, at the s e time, they pr serv a very h gh lateral resol inversion stra egie even if th latter approa hes are, in g eral, ore a depth of the targets and of retrieval f true resistivity values. Here, we di Network (NN) techniques, capable of reconstructing resistivity mode inversion strategy, but in a fraction of the time: seconds on a laptop vers demonstrate the advantages of the proposed novel approach on synthetic Methodologie . ATEM data re usually inverted by minimizing an o misfit term plus a stabilize . Hence, h objective functional to be mi imi P ( λ ) ( d , m ) = ‖ W d ( d obs − F ( m ) ) ‖ L 2 2 + λ where (a) d obs is the vector of the measurements; (b) m is the vector forward modelling operator connecting the model m to the correspondin of the process and the characteristics of the acquisiti n syste ; (d) based on the measur ment uncertainty; (e) s ( m ) is the stabilizer incor resistivity model to be recovered; (f) the multiplier λ controls the b information. In the deterministic s heme we are using here to asse approach based on NN, we consider a one-di ensional model paramet local 1D assumption. So, each individual data sounding, and each asso from the adjacent ones. More precisely, while the forward modelling constraint between the neighboring models is imposed by means of t i always one-dimensional, a ral constrain betwee the n ighbo ing models is imposed by means of the regularization term. In this respect, concerning the stabilizer choic , we adopt the very common option of ersion of Airborne TDEM data. S, Denmark mark ery quickly after, or ven during, the data collection would be crucial, adjusting the location of the proposed fl ght lines during n airbo ne d of readiness ould have a l rge impact in terms f maximization of rements to be acquired. Besides, the relevance of fast tools for EM data is demonstrated by the r utine use of Conductivity-Depth xploration. In fact, CDIs are extremely efficien from a computation l serve a very hig lateral r solution. Hence, they ar often pref rred to aches are, in general, more accurate in terms of reconstruction of the resistivity va ues. Here, we discuss a novel approach, based on Neural constructing re istivity models w th a quality c mparable with th time: seconds on a ptop versus hours on a computat onal server. We d novel approach on synthetic and field datasets (Bai t al., 2020). inverted by minimizing an objectiv functional nsisting of a data ective functional to be inimized is oft n w itten: , m ) = ‖ W d ( d obs − F ( m ) ) ‖ L 2 2 + λ s ( m ) , (1) rements; (b) m i the v ctor of the mod l param ters; (c) F is the model m t the corresp nding dat ; F takes into account t physics the acquisition system; (d) W d is the data weighting matrix usually ) s ( m ) is the stabilizer incorporating the prior knowledge about the multiplier λ controls the balance etween the data and the prior e we ar using here to assess the p rformances of the alternative e-dimensional model parameterization. Hence, m and F rely on the equal to the minimum gradient orm stabilizer ersion of Airborne TDEM data. S, Denmark nmark very quickly after, or even dur ng, the dat coll ction would be crucial r adjusting the location of the proposed flight l nes during an airborne ind of readiness could have a large impact in term of maximiz tion of urements to be acquired. Besides, the relevance of fast tools for TEM data is demonstrated by the routine use of Conductivity-Depth exploration. In fact, CDIs are extremely efficient from a computational reserve a very high lateral resolution. Hence, they are ften preferred to oaches are, in gener l, more accurat in terms of reconstruction of the resistivit values. Here, we discuss n v l approa h, based on Neural econstructing resistivity m del with a quality comparable with the time: s conds on a laptop versus hours on a c mputational server. We ed novel approach on synthetic and field datas ts (Bai e al., 2020). inverted by minimizing an objective fu ctional consisting f a data jective functional to b min mized is of en written: , m ) = ‖ W d ( d obs − F ( m ) ) ‖ L 2 2 + λ s ( m ) (1) uremen s; (b) m the vector of th mod l parameters; (c) F s the e model m to the corresponding data; F take into account the physics the acquisition system; (d) W d is the da weighting matrix usually e) s ( m ) is the stabilizer i corporating the pri r knowledge about the e mul iplier λ controls th balance between the dat an the prior e we are using here to s ess the performances of the alternative ne-dimensional model parameterization. Hence, m and F rely on the l data sounding, and each associat d model, is handled independently while the forward modelling F is always one-dimensional, a lateral els is imposed by means of the reg larization t rm. In this respec , p th very common option of s ( m ) equal o the minimum gradient ( m )= ‖ ∇ m ‖ L 2 2 . (2) bution is locally 1D, the stabilizer a ts both al ng the v rtical ( z ) and the level of laterally coherency (without be ng truly 2D/3D). This is the (2) Hence, even though the conductivity distribution is locally 1D, the stabilizer acts both along the vertic l ( a. t ollection would be crucial, flight lines during an airborne ct in terms of maximization of relevance of fast tools for ne use of Conductivity-Depth efficient from a computa ional c , hey are often preferred to terms of reconstruction of the vel approach, based on Neural quality comparable with th n a computation l serve . We datasets (Bai et al., 2020). nctional consisting of a data n w itt n: (1) od l parameters; (c) F i the takes into account the physics data weighting matrix usually he prior knowledge about the tween the data and the prior rformances of the alternati e Hence, m and F rely on the del, is handled independently ys one-dimensional, a lat ral ization t rm. In this respect, ual to the m imum gradie t (2) h alo g t i z ) and th ing truly 2D/3D). Th s is the t e horizontal ( Real-Time Neural Network Inversion of Airborne TDE P. Bai, Univ. of Cagliari G. Vignoli, U iv. of Caglia i, Italy & GEUS, Denmark A. Viezzoli, Aarhus Geophysics ApS, Denmark G. Vacca, Univ. of Cagliari, Italy The possibility of getting reliable results very quickly after, or even during, t not merely for quality check, but als for adjusting the location of the prop time-domain (ATEM) acquisition. This kind of readiness could have a large the Value of Information of the measurements to be acquired. Beside reconstr cting resistivity models from ATEM data is demonstrated by the Im ging (CDI) methodologies in mine al exploration. In fact, CDIs ar extre per pective and, at the same time, they pr se ve a v ry hi h lateral resoluti inv rsion stra egies even f the latter approaches ar , i general, more accur d pth of the target a d of retrieval of true resistivity valu s. Here, we discus Network (NN) techniques, capable o reconstru ting re i tivity mod ls w inversio str tegy, but in a fract on of the time: seconds on a l p op versus h demonstrat th advantages of the prop sed nov l approach on synthetic and Method logies. ATEM data are usually inverted by minimizing an object misfit term plus a stabilizer. Hence, th object ve functional to be m nimized P ( λ ) ( d , m ) = ‖ W d ( d obs − F ( m ) ) ‖ L 2 2 + λ s ( where ( ) d obs is the v ctor of the measurements; (b) m is the vector of t forward modelling operator connecting the model m to the corresponding da of the process nd the characteristics of the acquisition syst m; (d) W d is based o the measurem t uncertainty; (e) s ( m ) is the stabilizer incorpora resistivity odel to be recovered; (f) the multiplier λ controls the balan information. In the deterministic scheme we are using here to assess t approach based on NN, we consider a one-dimensional model parameteriz local 1D assumption. So, each i divid l data sounding, and each associat fro th adjacen ones. M re pr cisely, while the forward modelling F is constr int between the n ighboring models is i pos d by means of the r concerning the stabi izer choice, we adopt the very common option of s ( norm stabilizer s ( m )= ‖ ∇ m ‖ L 2 2 . Henc , even though the c ductivity distribution is locally 1D, the stabilizer ac horiz al x ) dir ction, enforcing some level of laterally coherency (witho irectio , enforci g some level of laterally coherency (w hout being truly 2D/3D). This is he essence of the Spatially Constrained Inversion (Viezzoli et al. 2009; Vign li et al. 2014; Ley-Cooper et al. 2015; Vignoli et al. 2017). I this framework, the value of of Airborn TDEM data. ark ly after, or even during, the data collection would e crucial, g the location of the proposed flight lines du ing a airborne din ss could have a large mpact in terms of maximizati of to be acquired. B sides, the r levance of fast tools for a is dem nstrated by the routine se of Co ductivity-D pth n. In fact, CDIs are ext emely efficien from a computational very high lateral resolution. Hence, they are often preferred to e, in general, more accurate in terms of reconstruction of the y values. Here, we discuss a novel approach, based on Neural ting resistivity models with quality co parabl with the onds on a laptop versus hours on a computational server. We pproach on synthetic and field datasets (Bai et al., 2020). by minimizi g an objective funct onal consisting f a data ctional to be minimized is often ritten: d ( d obs − F ( m ) ) ‖ L 2 2 + λ s ( m ) , (1) ; (b) m is the vector of the model parameters; (c) F is the to the corresponding data; F takes i to account th physics is chosen a poste iori in order to meet th following condition co cerning the chi-squared value